Resources » Help center » Crop best detected object to new image

Crop best detected object to new image

This task is used to crop the best object detected in an image, based on the various configurable criteria.

In the input parameters it is expected:

Input directory, where the images on which the detection is to be performed are located.
Output directory, where the result of the detections will be saved.
Path of the dmod file, which will be used as a model to perform the detections.
Confidence threshold: Accuracy of the detections performed by the AI model.
Only detections with a confidence higher than the one set in this parameter will be used for the rest of the processes.
Accuracy 0, implies that all detections will be evaluated, this may introduce erroneous detections in the process.
Accuracy very close to 1, implies that only the clearest detections for the model will be evaluated, this may bias the detections of valid objects.
By default, the value of 0.2 dramatically decreases anomalous detections and allows enough flexibility to maximize the number of detections.

In the advanced parameters we can configure:

Copy Images without detection: Copy to the output directory of the images in which no detection was performed.
Number of threads: The number of parallel threads of execution.
Packet processing size: The number of images that are processed in each parallel thread.
Buffer width: Percentage of horizontal enlargement on the detection performed with AI. The percentage is calculated as a function of the detection width. This is used to avoid marking the detection at the limit of the object and to give a margin.
Buffer height: Percentage of vertical magnification over the detection made with AI. The percentage is calculated based on the height of the detection. This is used to avoid marking the detection at the limit of the object and to give a margin.
Merge type: type of merging of overlapping objects of the same class, more details in the merge section.
Merge intersection: Percentage of intersection by which the detections will be merged, more details in the merge section.
Laplacian threshold: Quality threshold for an image to be processed.
This value allows you to filter out-of-focus images.
The default value is 100, it is appropriate to perform several tests to determine which value best suits the type of images evaluated.
If the value of this parameter is 0, the threshold is not applied.
The method we use to determine the quality of an image is the Laplacian operator, which allows us to obtain a mathematical parameter of image sharpness by studying the edges of objects. Thus a poorly focused image will have a lower Laplacian value than one in which the edges are clearly distinguishable.
Process by parts: This option allows processing the image by parts, intended for very large images.
Window Width: Measure in pixels of the width of the frame, when processing by parts.
Window Height: Measure in pixels of the height of the frame, when processing by parts.
Overlap value: Measurement in pixels, when processed in parts.
Area Weight: Weighting value to be used to determine the best object detected.
Score Weight: Weighting value that will be used to determine the best detected object.

Output example:

Blur detections in image

Draw detections in video