Grounding DINO Prelabeling

This page outlines the unit test coverage for the gdino_prelabelling.py script. The module runs object detection on a directory of images using the Grounding DINO model and saves prediction results in JSON format.

Coverage Overview

This test suite includes validation for:

  • File scanning via _get_image_files

  • End-to-end behavior of generate_gd_prelabelling

  • Device detection and fallback logic

  • Handling of corrupted or invalid image files

  • Handling of empty or missing directories

  • JSON output structure and formatting

Constants and Configs

The following are tested either directly or via mock configuration injection:

  • TEXT_PROMPTS

  • BOX_THRESHOLD

  • TEXT_THRESHOLD

  • Device resolution: "cuda", "cpu", "auto"

  • Paths to model weights and config files

Functions Tested

_get_image_files(directory)

  • Returns: Valid image paths (.jpg, .jpeg, .png)

  • Tests:

    • Valid image files

    • Mixed content (image + non-image)

    • Empty directories

    • Nonexistent directory (expected to raise error upstream)

generate_gd_prelabelling(...)

Core test focus:

  • Successful predictions written to JSON

  • Skipped or unreadable files are logged

  • Predictions include class, confidence, bounding box

  • All output files use expected format

  • Handles model loading errors

  • Applies config thresholds correctly

  • Automatically creates output directories if missing

  • Uses fallback device detection logic when set to "auto"

Key Edge Cases Tested

  • Empty folder: Returns no predictions and exits gracefully

  • Non-image files: Skipped without error

  • Corrupted images: Skipped and logged

  • Missing model file: Raises FileNotFoundError

  • Output folder does not exist: Auto-created

  • Multiple valid predictions per image: Verified in output structure

  • Verbose mode: Confirms detailed logs are printed to stdout

Example Test Assertions

  • Count of processed images matches input count

  • Prediction JSON contains required keys: bbox, confidence, class, source

  • Output filenames correspond to input image names

  • Invalid images do not result in written JSON

  • torch_device settings are passed to model correctly

Summary

This test suite ensures that gdino_prelabelling.py:

  • Works as expected across different environments (CPU, GPU, MPS)

  • Produces consistently formatted output

  • Handles missing, corrupted, or unexpected files gracefully

  • Is configurable via the pipeline-level config and can be extended

The use of mocks and temporary directories isolates test behavior from model internals, ensuring that unit-level functionality is verified in a reproducible way.