latest/how_to/filter_detections/

[giscus[bot]]

latest/how_to/filter_detections/

A set of easy-to-use utilities that will come in handy in any computer vision project.

https://supervision.roboflow.com/latest/how_to/filter_detections/

[Pencil77]

can i use it to filter the predictions in video?

[onuralpszr]

Yes you can filter videos as well.

[casperdias]

Hi can you help me I have this code:

from ultralytics import YOLO
import supervision as sv
import numpy as np
import cv2

# Replace 'rtsp://your_rtsp_stream_url' with your actual RTSP stream URL
rtsp_url = 'rtsp_url'

# Create a VideoCapture object
cap = cv2.VideoCapture(rtsp_url)
MODEL = "yolov8n.pt"
model = YOLO(MODEL)
model.fuse()

selected_classes = [0]

video_info_url = sv.VideoInfo(width=1920, height=1080, fps=15)
# settings
LINE_START = sv.Point(974, 757)
LINE_END = sv.Point(1450, 843)

# Check if the camera opened successfully
if not cap.isOpened():
    print("Error: Couldn't open the RTSP stream.")
    exit()

# create BYTETracker instance
byte_tracker = sv.ByteTrack(track_thresh=0.25, track_buffer=30, match_thresh=0.8, frame_rate=25)

# create VideoInfo instance
video_info = sv.VideoInfo.from_video_path(rtsp_url)

# create LineZone instance, it is previously called LineCounter class
triggering_anchors = [sv.Position.BOTTOM_LEFT, sv.Position.BOTTOM_RIGHT]
line_zone = sv.LineZone(start=LINE_START, end=LINE_END, triggering_anchors=triggering_anchors)

# create instance of BoxAnnotator
box_annotator = sv.BoxAnnotator(thickness=2, text_thickness=2, text_scale=1)

# create instance of TraceAnnotator
trace_annotator = sv.TraceAnnotator(thickness=2, trace_length=50)

# create LineZoneAnnotator instance, it is previously called LineCounterAnnotator class
line_zone_annotator = sv.LineZoneAnnotator(thickness=1, text_thickness=1, text_scale=0.5)

with sv.VideoSink(target_path='target_video.mp4', video_info=video_info_url) as sink:    
    while True:
        try:
            ret, frame_ori = cap.read()

            if not ret:
                print("Error: Couldn't read frame.")
                raise Exception("Error: Couldn't read frame.")

            results = model(frame_ori, verbose=False)[0]
            detections = sv.Detections.from_ultralytics(results)
            # only consider class id from selected_classes define above
            detections = detections[np.isin(detections.class_id, selected_classes)]
            # tracking detections
            detections = byte_tracker.update_with_detections(detections)
            labels = [
                f"#{tracker_id} {model.model.names[class_id]} {confidence:0.2f}"
                for confidence, class_id, tracker_id
                in zip(detections.confidence, detections.class_id, detections.tracker_id)
            ]
            annotated_frame = trace_annotator.annotate(
                scene=frame_ori.copy(),
                detections=detections
            )
            annotated_frame=box_annotator.annotate(
                scene=annotated_frame,
                detections=detections,
                labels=labels)

            # update line counter
            results_line = line_zone.trigger(detections)

            # Trigger Orang Masuk dan Keluar
            if results_line[0].size > 0:
                if results_line[0][0] == True:
                    print("Ada Orang Masuk")
                    with sv.ImageSink(target_dir_path="result") as sink:
                        for xyxy in detections.xyxy:
                            cropped_image = sv.crop_image(image=frame_ori, xyxy=xyxy)
                            sink.save_image(image=cropped_image)

            if results_line[1].size > 0:
                if results_line[1][0] == True:
                    print("Ada Orang Keluar")

            cv2.imshow('RTSP Stream', line_zone_annotator.annotate(annotated_frame, line_counter=line_zone))

            sink.write_frame(frame=line_zone_annotator.annotate(annotated_frame, line_counter=line_zone))

            if cv2.waitKey(1) & 0xFF == ord('q'):
                break

        except Exception as e:
            print("Error occurred:", e)
            cap = cv2.VideoCapture(rtsp_url)

# Release the VideoCapture and close the window
cap.release()
cv2.destroyAllWindows()

When someone enters, the code should cut the image and continue, but instead it creates an error occurred: 'ImageSink' object has no attribute 'write_frame' and makes it not respond.

Whats wrong with this code??

[kelvindbk]

The problem is in this line

with sv.ImageSink(target_dir_path="result") as sink:

this overrides the previous sink variable

with sv.VideoSink(target_path='target_video.mp4', video_info=video_info_url) as sink:

[codin-son]

hello, how can we filter using IOU? im using inferencepipeline for the detections, i want to make the iou can be modify by the user but it seem like the iou parameter for inferencepipeline init will not change when user modify the iou value

[SkalskiP]

Hi @codin-son 👋🏻 If you are unhappy with results coming straight from InferencePipeline try using sv.Detections.with_nms. This will allow you to use Non-Max Suppression - an algorithm that leverages Intersection over Union (IoU) to filter detections.

[morsechim]

Amazingly, it makes computer vision development much easier.

[EyadMas]

[EyadMas]

detections = detections[[4000 < detections.area < 18000]]

[EyadMas]

I want to remove detections that are so far from the camera and detections that are so close to the camera, do you find this way is correct or there is a better solution ?

[EyadMas]

I did fix it with np.logical_and(4000 < detections.area, detections.area< 18000), but kindly can you answer me is there a better way to remove objects far from the camera lens and too close to it from the detections (I guess filter is not the best way since it consumes more time ==> detetcion already has been done) so I am wonder if there is a way to make the detector remove them, or something like that.

[LinasKo]

This is the simplest way.

Other solutions are considerably more difficult as you need to distinguish between small objects, and objects that are far
away. This may involve depth estimation via depth cameras, depth-recovering models or triangulation.