Filters

import cv2
import pixelflow as pf
from ultralytics import YOLO

# Load model and run inference
model = YOLO("yolov8n.pt")
image = cv2.imread("traffic_scene.jpg")
outputs = model.predict(image)
detections = pf.detections.from_ultralytics(outputs)

# Basic filtering: keep high-confidence detections
high_conf = detections.filter_by_confidence(0.8)
print(f"High confidence: {len(high_conf)}/{len(detections)} detections")

# Conservative filtering for critical applications
critical = detections.filter_by_confidence(0.95)
print(f"Very high confidence: {len(critical)} detections")

# Method chaining with other filters
filtered = detections.filter_by_confidence(0.7).filter_by_class_id([0, 2])
print(f"High-conf people and cars: {len(filtered)} detections")

# Moderate filtering for analysis
moderate = detections.filter_by_confidence(0.5)
print(f"Moderate confidence: {len(moderate)} detections")

import cv2
import pixelflow as pf
from ultralytics import YOLO

# Load model and run inference
model = YOLO("yolov8n.pt")
image = cv2.imread("street_scene.jpg")
outputs = model.predict(image)
detections = pf.detections.from_ultralytics(outputs)

# Filter for specific class by numeric ID (COCO format)
people = detections.filter_by_class_id(0)  # person class
print(f"Found {len(people)} people")

# Filter for multiple vehicle classes
vehicles = detections.filter_by_class_id([2, 3, 5, 7])  # car, motorcycle, bus, truck
print(f"Found {len(vehicles)} vehicles")

# Filter by class name string
dogs = detections.filter_by_class_id("dog")
print(f"Found {len(dogs)} dogs")

# Mixed ID types for flexible filtering
targets = detections.filter_by_class_id(["person", 2, "bicycle"])
print(f"Found {len(targets)} people, cars, or bicycles")

# Consolidate vehicle classes to generic "vehicle"
vehicles = detections.remap_class_ids([2, 3, 5, 7], "vehicle")

# Remap multiple animal classes to "animal"
animals = detections.remap_class_ids(["dog", "cat", "bird"], 1)

# Single class remapping
motorcycles_as_bikes = detections.remap_class_ids(3, "bike")

# Remove very small noise detections
filtered = detections.filter_by_size(min_area=100)

# Remove very large detections (likely false positives)
filtered = detections.filter_by_size(max_area=50000)

# Keep medium-sized objects only
medium = detections.filter_by_size(min_area=1000, max_area=10000)

# Filter for tall, narrow objects (people)
people = detections.filter_by_dimensions(min_height=100, max_width=80)

# Remove extremely wide detections (likely errors)
cleaned = detections.filter_by_dimensions(max_width=500)

# Keep medium-sized rectangular objects
medium = detections.filter_by_dimensions(

# Keep square-ish objects only
squares = detections.filter_by_aspect_ratio(min_ratio=0.8, max_ratio=1.2)

# Keep only wide objects (banners, signs)
wide_objects = detections.filter_by_aspect_ratio(min_ratio=2.0)

# Keep only tall objects (people, poles)
tall_objects = detections.filter_by_aspect_ratio(max_ratio=0.5)

# Filter for typical car aspect ratios
cars = detections.filter_by_aspect_ratio(min_ratio=1.5, max_ratio=2.5)

# Setup zones and update detections
zone_manager = pf.zones.ZoneManager()
detections = detections.update_zones(zone_manager)

# Keep only detections in parking areas
parking = detections.filter_by_zones(["parking_1", "parking_2"])

# Exclude detections from restricted areas
allowed = detections.filter_by_zones(["restricted", "private"], exclude=True)

# Filter by numeric zone IDs
zone_1_only = detections.filter_by_zones(1)

# Keep only detections in center 80% of frame
center_detections = detections.filter_by_position(

# Remove tiny noise detections (< 0.01% of frame)
cleaned = detections.filter_by_relative_size(

# Keep only persistent objects (tracked for at least 5 seconds)
persistent = detections.filter_by_tracking_duration(min_seconds=5.0)

# Find short-term detections (tracked 2-10 seconds)
short_term = detections.filter_by_tracking_duration(

# Keep only objects first detected in early frames (0-100)
early_objects = detections.filter_by_first_seen_time(start_time=0, end_time=100)

# Focus on objects that appeared recently (after frame 500)
recent_objects = detections.filter_by_first_seen_time(start_time=500)

# Analyze objects from specific time window
window_objects = detections.filter_by_first_seen_time(

# Keep only successfully tracked objects
tracked_only = detections.filter_tracked_objects(require_tracker_id=True)

# Get newly detected objects (not yet tracked)
new_detections = detections.filter_tracked_objects(require_tracker_id=False)

# Analyze tracking success rate
total = len(detections)
tracked = len(detections.filter_tracked_objects(True))
print(f"Tracking rate: {tracked/total:.1%}")

# Remove highly overlapping detections (conservative)
cleaned = detections.remove_duplicates(iou_threshold=0.8, keep='first')

# Aggressive duplicate removal, keep best confidence
best_only = detections.remove_duplicates(

# Find detections that overlap significantly with others
overlapping = detections.filter_overlapping(min_overlap=0.3)

# Find objects that overlap with people (co-occurrence analysis)
near_people = detections.filter_overlapping(