Skip to main content
The classification pipeline processes each video frame through segmentation and visualization stages to detect and classify trash objects in real-time.

​
Pipeline Overview

The TrashClassificator class in processor.py:7 orchestrates the entire classification workflow through its frame_processing() method. 
from trash_classificator.processor import TrashClassificator

classificator = TrashClassificator()
processed_image, status = classificator.frame_processing(image)

​
Processing Steps

1

Initialize Components

The classifier initializes two main components:
self.segmentation = SegmentationModel()
self.draw_detections = Drawing()
  • SegmentationModel: Handles trash detection and tracking
  • Drawing: Manages visualization of detections
2

Trash Segmentation

The input image is copied and passed to the segmentation model:
trash_image = image.copy()
trash_track, trash_classes, device = self.segmentation.inference(trash_image)
Returns:
  • trash_track: Detection results with boxes, masks, and tracking IDs
  • trash_classes: Dictionary mapping class IDs to names
  • device: PyTorch device used for inference
See Segmentation Model for details.
3

Detection Validation

The pipeline checks if any tracked objects were detected:
for trash in trash_track:
    if trash.boxes.id is None:
        return image, 'No trash detected'
If no boxes have tracking IDs, the original unmodified image is returned.
4

Draw Detections

For each detected object, visualizations are added:
image_draw = image.copy()
image_draw = self.draw_detections.draw(image_draw, trash, trash_classes, device)
return image_draw, 'Trash detected'
This includes masks, bounding boxes, class labels, and tracking paths.

​
Input/Output Specifications

def frame_processing(self, image: np.ndarray):
    """
    Args:
        image: NumPy array in BGR format (OpenCV convention)
               Shape: (height, width, 3)
               Dtype: uint8
    
    Returns:
        Tuple[np.ndarray, str]:
            - Processed image with visualizations
            - Status message: 'Trash detected' or 'No trash detected'
    """

​
Processing Flow Diagram

​
Error Handling

The pipeline implements graceful degradation:
No Detection HandlingWhen trash.boxes.id is None, the pipeline returns the original image unchanged with status 'No trash detected'. This prevents visualization errors when no objects are tracked.

​
Common Issues

IssueCauseSolution
Empty tracking IDsNo objects meet confidence thresholdLower conf parameter in segmentation
Processing lagHigh resolution inputReduce imgsz parameter or input size
Memory errorsLarge batch of framesProcess frames individually

​
Performance Considerations

The pipeline creates multiple image copies to preserve the original:
trash_image = image.copy()  # For segmentation
image_draw = image.copy()   # For drawing
This adds memory overhead but ensures the input remains unchanged.

​
Configuration Parameters

The segmentation inference in segmentation/main.py:24 accepts these parameters: 
results = model.track(
    image, 
    conf=0.55,        # Confidence threshold (0-1)
    verbose=False,    # Suppress console output
    persist=True,     # Maintain tracking across frames
    imgsz=640,        # Input image size
    stream=True       # Enable streaming mode
)

​
Parameter Tuning

  • conf: Lower values detect more objects but increase false positives
  • imgsz: Larger sizes improve accuracy but reduce speed
  • persist: Must be True for consistent object tracking

​
Next Steps

Video Processing

Learn how to process video streams in real-time

Drawing Detections

Customize visualization of detected trash

Build docs developers (and LLMs) love

Get started for freeTalk to us