Skip to main content

Changelog

DriveID filter release notes

v0.1.8 - 2025-09-27​

Changed​

  • Updated documentation

v0.1.7 - 2025-08-06​

Modified​

  • Updated dependencies

v0.1.6 - 2025-08-06​

Modified​

  • Support for model context file name
  • Updated dependencies

v0.1.5 - 2025-07-30​

Added​

  • Support for model info-context

v0.1.4 - 2025-07-15​

Changed​

  • Updated dependencies

v0.1.3 - 2025-05-22​

Changed​

  • Updated dependencies

v0.1.2 - 2025-05-22​

Added​

  • Initial release of the DriveID filter using a custom-trained Faster R-CNN model.
  • Supports detection of license plates in image frames with the following features:
    • Loads a Torch model from a configurable model_path
    • Detects plates and returns bounding boxes with confidence scores
  • Frame-level control:
    • Skips processing for frames with metadata flag skip_plate_detection: true
  • Confidence threshold:
    • Discards detections below a configurable confidence_threshold (default: 0.7)
  • Output formatting:
    • Writes detection results to a configurable output_json_path (if write_detections_to_json is enabled)
    • Each record includes:
      • frame_id
      • List of detected plates with bounding box and confidence score
  • Forwarding support:
    • Optionally forwards polygon ROIs to downstream consumers via frame.data['meta'][roi_output_label]
    • Enabled using forward_detection_rois
    • Configurable label name via roi_output_label
  • Debug mode:
    • Enables verbose logging when debug is true
  • Device auto-detection:
    • Automatically uses CUDA if available; otherwise falls back to CPU
  • Environment variable configuration:
    • All config fields can be overridden via FILTER_* env vars (e.g., FILTER_MODEL_PATH, FILTER_DEBUG)
  • Includes PIL and OpenCV-based preprocessing with TorchVision transforms

Changed​

  • Scaled predicted bounding boxes from model coordinates to original frame dimensions using width/height ratio
  • Improved transform pipeline for input normalization
  • Adjusted logging to show frame-wise outputs and polygon forwarding activity

Fixed​

  • Fixed potential mismatch between input frame resolution and model preprocessing dimensions
  • Ensured bounding box coordinates are properly rounded and cast to integers
  • Resolved potential file writing issues by safely creating output directories for JSON logging

Internal​

  • Refactored model loading into a separate method with dynamic class predictor injection
  • Consolidated TorchVision transform logic into a reusable get_transform() method
  • Enhanced logging throughout setup, process, and shutdown phases

Experimental​

  • Polygon ROI forwarding via rectangular box conversion for downstream processing