Filter Detections Error about supervision

[XiaJunfly]

@rolson24 @SkalskiP
Thank you for your help. It works. However，the ID of each object within the set zone is chaotic，as IDs are assigned throughout the entire map range. As you say，the object is inactivate（se here （and I don‘t understand why this happend），so the tracker_ids is empty. So，I used this code to deal with this problems：

import time
from tqdm import tqdm
import cv2
import numpy as np
from ultralytics import YOLO
import supervision as sv
import matplotlib.pyplot as plt

model = YOLO('yolov8x.pt')

VIDEO_PATH = 'work_people.mp4'
# VIDEO_PATH = 'traffic_analysis.mov'

video_info = sv.VideoInfo.from_video_path(VIDEO_PATH)

polygons = [
   np.array([[100, 600], [600, 485], [480, 390], [80, 450]]),
   np.array([[950, 400], [1200, 340], [1100, 280], [800, 330]])
]
# polygons = [
#     np.array([[100, 600], [600, 485], [480, 390], [80, 450]])]

# polygons = [
#     np.array([[592, 282], [900, 282], [900, 82], [592, 82]]),
#     np.array([[1250, 282], [1250, 530], [1450, 530], [1450, 282]])
# ]

zones = [sv.PolygonZone(polygon=polygon) for polygon in polygons]
colors = sv.ColorPalette.DEFAULT
zone_annotators = [sv.PolygonZoneAnnotator(zone=zone, color=colors.by_idx(index), thickness=2)
                  for index, zone in enumerate(zones)]
box_annotators = [sv.BoundingBoxAnnotator(color=colors.by_idx(index), thickness=2)
                 for index in range(len(polygons))]
track1 = sv.ByteTrack()
track2 = sv.ByteTrack()
trackers = [track1, track2]
label_annotators = [sv.LabelAnnotator(color=colors.by_idx(index)) for index in range(len(polygons))]
trace_annotators = [sv.TraceAnnotator(color=colors.by_idx(index), trace_length=50) for index in range(len(polygons))]

# print('len list', len(label_annotators), label_annotators[0])


def process_frame(frame: np.ndarray, i) -> np.ndarray:
   results = model(frame, imgsz=1280, verbose=False, show=False, device='cuda:0')[0]

   # supervision
   detections = sv.Detections.from_ultralytics(results)

   annotated_frame = frame.copy()
   i = 0
   for zone, zone_annotator, box_annotator, label_annotator, trace_annotator, tracker in (
           zip(zones, zone_annotators, box_annotators, label_annotators, trace_annotators, trackers)):
       # print('i', i)

       mask = zone.trigger(detections=detections)
       # print("mask", mask)

       # get target
       detections_filtered = detections[mask]
       # print('dection-get', detections_filtered)

       detections_filtered = tracker.update_with_detections(detections_filtered)
       # print('detections_filtered.tracker_id', detections_filtered.class_id, detections_filtered.tracker_id)

       labels = [
           f"#{tracker_id} {results.names[class_id]}"
           for class_id, tracker_id
           in zip(detections_filtered.class_id, detections_filtered.tracker_id)
       ]
       # print('label', labels, len(labels))

       # box
       # annotated_frame = box_annotator.annotate(annotated_frame, detections=detections_filtered)
       # cv2.imshow("box", annotated_frame)
       # cv2.waitKey(0)
       #
       # annotated_frame = label_annotator.annotate(annotated_frame, detections=detections_filtered, labels=labels)
       # # cv2.imshow("ID", annotated_frame)
       # # cv2.waitKey(0)
       #
       # # track
       # annotated_frame = trace_annotator.annotate(annotated_frame, detections=detections_filtered)
       # # cv2.imshow("trac", annotated_frame)
       # # cv2.waitKey(0)
       #
       # annotated_frame = zone_annotator.annotate(scene=annotated_frame)

       if len(labels) >= 1:
           # ID and labels
           annotated_frame = label_annotator.annotate(annotated_frame, detections=detections_filtered, labels=labels)

           # track
           annotated_frame = trace_annotator.annotate(annotated_frame, detections=detections_filtered)


           annotated_frame = zone_annotator.annotate(scene=annotated_frame)

       else:
           annotated_frame = zone_annotator.annotate(scene=annotated_frame, )

       i += 1

   pbar.update(1)

   return annotated_frame


filehead = VIDEO_PATH.split('/')[-1]
OUT_PATH = "out-" + filehead

with tqdm(total=video_info.total_frames - 1) as pbar:
   sv.process_video(source_path=VIDEO_PATH, target_path=OUT_PATH, callback=process_frame)

[SkalskiP]

Hi @XiaJunfly 👋🏻 Try running this code:

  import time
  from tqdm import tqdm
  import cv2
  import numpy as np
  from ultralytics import YOLO
  import supervision as sv
  import matplotlib.pyplot as plt
  
  model = YOLO('yolov8x.pt')
  
  VIDEO_PATH = 'people-walking.mp4'
  
  video_info = sv.VideoInfo.from_video_path(VIDEO_PATH)
  
  polygons = [
      np.array([[100, 800], [500, 800],[500, 100], [100, 100]]),
      np.array([[600, 800],[800, 800], [800, 600], [600, 600]])
  ]
  
  zones = [sv.PolygonZone(polygon=polygon, frame_resolution_wh=video_info.resolution_wh) for polygon in polygons]
  colors = sv.ColorPalette.default()
  zone_annotators = [sv.PolygonZoneAnnotator(zone=zone, color=colors.by_idx(index), thickness=6)
                     for index, zone in enumerate(zones)]
  box_annotators = [sv.BoundingBoxAnnotator(color=colors.by_idx(index), thickness=2)
                    for index in range(len(polygons))]
  tracker = sv.ByteTrack()
  label_annotators = [sv.LabelAnnotator(color=colors.by_idx(index)) for index in range(len(polygons))]
  trace_annotators = [sv.TraceAnnotator(color=colors.by_idx(index)) for index in range(len(polygons))]
  
  def process_frame(frame: np.ndarray, i) -> np.ndarray:
  
      results = model(frame, imgsz=1280, verbose=False, show=False, device='cuda:0')[0]
  
  
      # supervision 
      detections = sv.Detections.from_ultralytics(results)
  
      annotated_frame = frame.copy()
      for zone, zone_annotator, box_annotator, label_annotator,  trace_annotator in (
              zip(zones, zone_annotators, box_annotators, label_annotators, trace_annotators)):
          print('i', zone, zone_annotator, box_annotator, label_annotator,  trace_annotator)
  
          mask = zone.trigger(detections=detections)
  
          # get target
          detections_filtered = detections[mask]
          print('dection', detections_filtered)
  
  
          detections_filtered = tracker.update_with_detections(detections_filtered)
          print('detections_filtered.tracker_id', detections_filtered.class_id, detections_filtered.tracker_id)
  
          labels = [
              f"#{tracker_id} {results.names[class_id]}"
              for class_id, tracker_id
              in zip(detections_filtered.class_id, detections_filtered.tracker_id)
          ]
          print('label', labels)
  
          # box
          annotated_frame = box_annotator.annotate(annotated_frame, detections=detections_filtered)
  
          # ID and labels
          annotated_frame = label_annotator.annotate(annotated_frame, detections=detections_filtered, labels=labels)
  
          # track
          annotated_frame = trace_annotator.annotate(annotated_frame, detections=detections_filtered)
          cv2.imshow("trac", frame)
          cv2.waitKey(0)
  
          annotated_frame = zone_annotator.annotate(scene=annotated_frame)
  
      pbar.update(1)
  
      return annotated_frame
  
  filehead = VIDEO_PATH.split('/')[-1]
  OUT_PATH = "out-" + filehead
  
  with tqdm(total=video_info.total_frames - 1) as pbar:
      sv.process_video(source_path=VIDEO_PATH, target_path=OUT_PATH, callback=process_frame)

[XiaJunfly]

@skylargivens
Thank you for your reply. I tried your code,but it shows the same error.

[SkalskiP]

@XiaJunfly can you reproduce this bug in google colab?

[rolson24]

Hi @SkalskiP and @XiaJunfly
I know what the issue is.

In this PR #1035 I changed the tracker to return the original detections only if they were tracked. It worked in all of my tests, except I forgot to test when none of the detections were tracked. Later after we merged the PR I noticed some bugs with the PR that when none of the detections are tracked, the function returns the original detections, but with the tracker_ids empty (see here). I fixed it in my PR here: #1076, but that was not included in 0.20.0.

I realize now that I should have fixed the issue in a separate PR for transparency, and I'm sorry about that.

For now, one solution could be to check if the tracker_ids are empty and if the they are just throw them away (hacky).

Here is a colab notebook with the broken code and the possible fix.

[SkalskiP]

Hi @XiaJunfly 👋🏻 I am very sorry for the confusion. As @rolson24 has already explained, we are experiencing some issues with the ByteTrack version released in supervision=0.20.0. I apologize once again. For this reason, I encourage you to use the pre-released version supervision=0.21.0rc3.

https://pypi.org/project/supervision/0.21.0rc3/

[SkalskiP]

I'm closing this issue. Feel free to reopen it if you have more questions.

btw, @rolson24 you are becoming our tracking expert! 🙌🏻