Inter-class adaptive threshold structure for object detection
Provided herein are systems and methods for applying adaptive classes thresholds to enhance object detection Machine Learning (ML) models by receiving a plurality of labeled feature vectors extracted from a plurality of images associated with a plurality of objects, one or more subsets of the plurality of feature vectors are associated with respective object(s) and labeled accordingly, computing an adaptive threshold for each object in a plurality of iterations, each iteration comprising: (1) computing deviation of a respective feature vector of the subset from an aggregated feature vector, (2) computing, in case the deviation is within a predefined value, a threshold enclosing the respective feature vector, and (3) adjusting the adaptive threshold to enclose the threshold of the respective feature vector and outputting the adaptive threshold(s) for classifying unlabeled feature vectors to class(s) of respective object(s) associated with the adaptive threshold(s) in which the unlabeled feature vectors fall.
1. A computer implemented method of applying adaptive classes thresholds to enhance object detection Machine Learning (ML) models, comprising:
receiving a plurality of labeled feature vectors extracted from a plurality of images associated with a plurality of objects, at least one subset of the plurality of feature vectors is associated with a respective one of the plurality of objects and labeled accordingly;
computing an adaptive threshold for the respective object in a plurality of iterations each for a respective one of the feature vectors of the subset, each iteration comprising:
computing a deviation of the respective feature vector from an aggregated feature vector aggregating all previously processed feature vectors of the subset,
computing, in case the deviation is within a predefined value, a threshold enclosing the respective feature vector by applying at least one ML model to compute similarity with each of the plurality of feature vectors not associated with the respective object, and
adjusting the adaptive threshold to enclose the threshold of the respective feature vector; and
outputting the adaptive threshold for classifying at least one unlabeled feature vector extracted by at least one ML model from at least one query image to a class of the respective object in case the at least one unlabeled feature vector falls within the adaptive threshold of the respective object.
2. The computer implemented method of claim 1 , wherein the plurality of objects are faces of people such that the at least one ML model is configured and trained for face recognition.
3. The computer implemented method of claim 1 , wherein the at least one ML model is a neural network.
4. The computer implemented method of claim 1 , wherein the aggregated feature vector is a common center computed based on the previously processed feature vectors of the subset and the deviation is computed based on a distance of the respective feature vector from the common center.
5. The computer implemented method of claim 1 , wherein the threshold and the adaptive threshold define respective multi-dimensional spatial regions having a plurality of dimensions corresponding to the number of features in the plurality of feature vectors.
6. The computer implemented method of claim 1 , wherein adjusting the adaptive threshold to enclose the threshold of the respective feature vector comprising adjusting a border line of the multi-dimensional spatial region defining the adaptive threshold to follow the border line of the multi-dimensional spatial region defining the threshold computed for the respective feature vector.
7. The computer implemented method of claim 6 , further comprising smoothing the border line of the multi-dimensional spatial region defining the adaptive threshold.
8. The computer implemented method of claim 1 , wherein a respective adaptive threshold is individually adjusted for each of the plurality of objects by adjusting the respective adaptive threshold according to the threshold computed for each feature vector of the plurality of features vectors associated with the respective object having a deviation from a respective aggregated feature vector which is within the predefined value.
9. A system for applying adaptive classes thresholds to enhance object detection Machine Learning (ML) models, comprising:
at least one processor executing a code, the code comprising:
code instructions to receive a plurality of labeled feature vectors extracted from a plurality of images associated with a plurality of objects, at least one subset of the plurality of feature vectors is associated with a respective one of the plurality of objects and labeled accordingly:
code instructions to compute an adaptive threshold for the respective object in a plurality of iterations each conducted for a respective one of the feature vectors of the subset, each iteration comprising:
computing a deviation of the respective feature vector from an aggregated feature vector aggregating all previously processed feature vectors of the subset,
computing, in case the deviation is within a predefined value, a threshold enclosing the respective feature vector by applying at least one ML model to compute similarity with each of the plurality of feature vectors not associated with the respective object, and
adjusting the adaptive threshold to enclose the threshold of the respective feature vector; and
code instructions to output the adaptive threshold for classifying at least one unlabeled feature vector extracted by at least one ML model from at least one query image to a class of the respective object in case the at least one unlabeled feature vector falls within the adaptive threshold of the respective object.