Skip to main content

​
Overview

While the Reciclaje AI model provides reliable waste detection for common scenarios, it’s important to understand its limitations. This page documents known edge cases and situations where the model may produce inaccurate or unreliable results.
Important: Understanding these limitations is crucial for deploying the model in production environments and setting appropriate expectations.

​
Environmental Limitations

​
Adverse Lighting Conditions

The model may fail or produce low-confidence predictions in challenging lighting scenarios:
Impact: Reduced detection accuracy or missed detectionsExamples:
  • Poorly lit recycling areas
  • Evening or nighttime outdoor conditions
  • Shadowed areas with insufficient ambient light
Mitigation:
  • Ensure adequate lighting in deployment areas
  • Use supplementary LED lighting for cameras
  • Consider adjusting camera exposure settings
Impact: Washed out features, difficulty distinguishing materialsExamples:
  • Direct sunlight on reflective surfaces
  • Strong overhead lighting causing glare
  • Bright backgrounds that reduce contrast
Mitigation:
  • Position cameras to avoid direct light sources
  • Use polarizing filters on camera lenses
  • Adjust camera settings to handle high dynamic range
Impact: Uneven detection performance across the frameExamples:
  • Partially shadowed objects
  • Transitional lighting (doorways, windows)
  • Flickering or unstable light sources
Mitigation:
  • Provide consistent, uniform lighting
  • Use diffused lighting to reduce shadows

​
Image Quality Issues

​
Noise and Artifacts

The model may fail when images contain significant noise or artifacts.
Problematic scenarios include:
  • High ISO Noise: Grainy images from low-light camera settings
  • Motion Blur: Fast-moving objects or camera shake
  • Compression Artifacts: Heavily compressed video streams
  • Lens Distortion: Wide-angle lenses causing warping
  • Out of Focus: Blurry or unfocused images
# Example: Low quality frame detection
if frame_quality_score < THRESHOLD:
    print("Warning: Low quality frame detected")
    # Consider skipping frame or requesting better image

​
Perspective and Orientation

​
Unusual Viewing Angles

The model is optimized for typical viewing perspectives:

Optimal Angles

  • Front-facing views
  • Slight angle variations (±30°)
  • Standard camera heights
  • Horizontal orientations

Challenging Angles

  • Top-down (bird’s eye) views
  • Extreme side angles
  • Upside-down orientations
  • Very close or far distances

​
Occlusion

Partially hidden or overlapping objects may cause:
  • Missed Detections: Object not recognized at all
  • Misclassification: Visible portion resembles different class
  • Multiple Boxes: Single object detected as multiple items
The model performs best when objects are clearly visible with minimal overlap.

​
Object-Specific Limitations

​
Rare or Uncommon Items

Critical Limitation: The model may not recognize waste items that were rarely observed or not included in the training dataset.
Problematic items may include:
  • Uncommon packaging designs
  • Regional or specialized products
  • Damaged items with unusual appearance
  • New materials not present during training
  • Mixed-material composite items

​
Size Extremes

Detection accuracy may decrease for:
  • Very Small Objects: Items occupying less than 5% of frame area
  • Very Large Objects: Items filling most of the frame
  • Distant Objects: Objects far from the camera
# Calculate object size relative to frame
box_area = (x2 - x1) * (y2 - y1)
frame_area = frame.shape[0] * frame.shape[1]
relative_size = box_area / frame_area

if relative_size < 0.05 or relative_size > 0.8:
    # Potential detection issues
    pass

​
Material-Specific Challenges

​
Transparent Materials

Glass detection can be particularly challenging:
  • Clear glass bottles may blend with backgrounds
  • Reflections can obscure true appearance
  • Transparency makes edge detection difficult

​
Reflective Surfaces

Metal items with high reflectivity:
  • Reflections can confuse the model
  • Shiny surfaces may appear different under various lighting
  • Chrome or polished metals more difficult than matte finishes

​
Similar Appearances

Some materials may be confused:
  • Metal cans vs. metallic plastic packaging
  • Paper cartons with plastic coating
  • Composite materials (e.g., Tetra Pak)

​
Confidence Score Considerations

​
Low Confidence Predictions

When the model is uncertain: 
conf = box.conf[0]

if conf < 0.5:  # 50% confidence threshold
    # Low confidence - use with caution
    print(f"Low confidence: {conf*100:.1f}%")
Consider implementing confidence thresholds in your application to filter out unreliable predictions.

​
False Positives and Negatives

  • False Positives: Non-waste items incorrectly classified as waste
  • False Negatives: Waste items not detected at all
Both are more likely under the limiting conditions described above.

​
Deployment Recommendations

​
Best Practices

To minimize limitations in production:
1

Environment Control

Optimize lighting and camera positioning for consistent conditions.
2

Quality Checks

Implement image quality validation before running inference.
3

Confidence Thresholds

Set appropriate confidence score thresholds for your use case.
4

Human Review

Include human verification for low-confidence or critical decisions.
5

Monitoring

Track model performance and detection patterns in production.

​
When to Use Human Verification

Consider requiring human review when:
  • Confidence scores are below threshold (e.g., less than 70%)
  • Multiple classes detected with similar confidence
  • Critical sorting decisions (e.g., medical waste)
  • Unusual or rare items encountered

​
Future Improvements

Potential areas for model enhancement:
  • Expanded Training Data: More diverse examples of edge cases
  • Data Augmentation: Better handling of lighting and perspective variations
  • Multi-Model Ensemble: Combining multiple models for robustness
  • Active Learning: Continuous improvement from deployment feedback

​
Next Steps

Model Overview

Review model specifications

Training Data

Learn about the training dataset

Quick Start

Start using the model

Build docs developers (and LLMs) love

Get started for freeTalk to us