Getting just one box for each object (Object Detection)

[lcoandrade]

I'm trying to make object detection predictions using my trained model:

bundle_uri = os.path.join(OUTPUT_DIR, 'model-bundle.zip')
predictlearner = ObjectDetectionLearner.from_model_bundle(
    model_bundle_uri=bundle_uri,
    training=False,
    output_dir=OUTPUT_DIR,
)
predictlearner.log_data_stats()

To make the predictions, I'm using this dataset (a geotif image):

prediction_dataset = ObjectDetectionSlidingWindowGeoDataset.from_uris(
        class_config=class_config,
        image_uri='/kaggle/input/objectdetectiontestimage/test_image.tif',
        size=IMG_SIZE,
        stride=STRIDE,
    )

And this code:

predictions = predictlearner.predict_dataset(
    dataset=prediction_dataset,
    raw_out=False,# raw_out False because I want class IDs
    numpy_out=True,
    predict_kw=dict(out_shape=IMG_SIZE),
    progress_bar=True,
)

pred_labels = ObjectDetectionLabels.from_predictions(
    windows=prediction_dataset.windows,
    predictions=predictions,
)

# Saving the predictions in a geojson file
predictions_path = os.path.join(OUTPUT_DIR, 'predictions.geojson')
pred_labels.save(
    uri=predictions_path,
    crs_transformer=prediction_dataset.scene.raster_source.crs_transformer,
    class_config=class_config,
)

The problem I wat to solve is the multiplicity of boxes in my predictions like we can see here:

How to output just one box?

[AdeelH]

ObjectDetectionLabels.prune_duplicates is what you want.

ObjectDetectionLabels.prune_duplicates(pred_labels, score_thresh, merge_thresh)