System and method for reducing false positives in object detection frameworks

A method for improving performance of an object detection and classification model comprising:

• obtaining an image annotated by the model having bounding boxes enclosing detected objects in the image;
• for each bounding box in the image:
• obtaining, from the model, a feature representation of the object enclosed in the bounding box;
• searching feature representations stored in a false positives gallery for a match with the feature representation of the object enclosed in the bounding box;
• displaying the result when no matches are found between the feature representation of the object and feature representations stored in the false positives gallery;
• receiving, from a user, an indication that the display result does not match the object;
• and
• storing the feature representation of the object in the false positives gallery.

A system comprising:

• an object detector configured to detect objects in an image;
• a feature extractor configured to extract feature representation of the detected objects;
• a classifier configured to classify the feature representations;
• and
• a false positives gallery, containing feature representations of had been previously classified incorrectly;
• wherein each feature representation is matched against feature representations contained in the false positives gallery to determine if the feature representation had previously been classified incorrectly.