Box Filtering

No addon config required

This add on computes the Inter Quartile Range on bounding box areas after the model inference (post_process) and filters out boxes that fall in upper bound.

IQR = Quartile3 – Quartile1 (Inter Quartile Range)
Lower Bound: (Quartile1 - 1.5 * IQR)
Upper Bound: (Quartile3 + 1.5 * IQR)

This addon has been implemented as workaround to object detectors that were trained on data that annotate multiple objects with one big bounding box (incorrect annotations).

Debug

Example of object initialization and post_process execution:

from vsdkx.addon.box_filtering.processor import BigBoxFilteringProcessor
addon_on_config = {
  'class': 'vsdkx.addon.box_filtering.processor.BigBoxFilteringProcessor'
  }
model_config = {
    'classes_len': 1, 
    'filter_class_ids': [0], 
    'input_shape': [640, 640], 
    'model_path': 'vsdkx/weights/ppl_detection_retrain_training_2.pt'
    }
    
 model_settings = {
    'conf_thresh': 0.5, 
    'device': 'cpu', 
    'iou_thresh': 0.4
    }
box_filter_processor = BigBoxFilteringProcessor(addon_on_config, model_settings, model_config)
#post_process execution 
 addon_object = AddonObject(
    frame=np.array(RGB image), #Required RGB image in numpy format
    inference=dict{
                boxes=[array([2007,  608, 3322, 2140]), array([ 348,  348, 2190, 2145])], 
                classes=[array([0], dtype=object), array([0], dtype=object)], 
                scores=[array([0.799637496471405], dtype=object), array([0.6711544394493103], dtype=object)], 
                extra={}}, 
    shared={}
    )
 addon_object = box_filter_processor.post_process(addon_object)

This step updates the addon_object.inference.boxes, addon_object.inference.scores and addon_object.inference.classes with the filtered bounding boxes, scores and classes.