Interactive system and method for improved collection and annotation of machine learning datasets

1 . A method for curation of training data for training machine learning models comprising:

capturing, using an image capture device, a reference image frame comprising an object of interest;

obtaining a reference datapoint for a label or a set of labels from the reference image frame; wherein the label or set of labels is associated with the object of interest;

annotating the reference datapoint to ascertain an expected annotation for each label or the set of labels; wherein the expected annotation comprises a geometrical shape;

establishing the expected annotations and a pre-defined quantity target as seed data for each label or the set of labels;

scaling the geometrical shape of the expected annotation based on a predetermined scaling factor;

obtaining a position of the scaled geometric shape as an annotation guideline;

shifting the position of the annotation guideline by varying the position in a two-dimensional coordinate system and displaying the shifted annotation guideline and a live image on the screen of the image capture device; wherein the live image comprises an image including the object of interest;

prompting a user to adjust the live image frame such that the object of interest is tightly encompassed inside the scaled and shifted annotation guideline displayed on the live image screen;

capturing a second live image frame and comparing the objects of interest in the captured second live image frame with the objects of interest in the reference image frame to determine a similarity score;

classifying, in real-time, the captured second live image frame as accepted, if the similarity score is above a pre-defined threshold;

storing the accepted image frames as training dataset for machine learning models; and

repeating, for newly captured live image frame, the steps of shifting, prompting, capturing, classifying and storing until the pre-defined quantity target criteria is satisfied, based on a comparison between the newly captured live image frame and the reference image frame.

2 . The method of claim 1 , wherein the geometrical shape comprises a rectangular bounding box, a square bounding box or a higher dimension polygonal bounding box.

3 . The method of claim 1 , wherein the reference image frame comprises one or more objects.

4 . The method of claim 3 , wherein each of the one or more objects is associated with a label from the set of labels.

5 . The method of claim 1 , wherein the pre-defined quantity target is a threshold number of captured image frames classified as accepted for each of the labels or the set of labels.

6 . The method of claim 1 , wherein predetermined scaling factor is a value by which the size of the geometrical shape of the expected annotation is scaled-up and/or scaled-down to obtain the size of the annotation guideline.

7 . The method of claim 1 , wherein shifting the position of annotation guideline comprises changing the two-dimensional co-ordinates of the annotation guideline using a shifting formula.

8 . The method of claim 1 , wherein prompting the user comprises providing a visual or textual indication.

9 . The method of claim 8 , wherein prompting the user further comprises checking for motion or blur, improper framing of the live image screen.

10 . The method of claim 1 , wherein similarity score is determined based on one or more of a blur detection, a noise detection, an improper framing detection and environmental conditions.

11 . The method of claim 10 , wherein classifying in real-time the captured images comprises, comparing the similarity score of the captured images to one or more of a benchmark criteria.

12 . The method of claim 11 , wherein the benchmark criteria is associated with a label or set of labels.

13 . The method of claim 1 , wherein the training dataset comprises audio dataset, image dataset, video dataset and/or textual dataset.

14 . The method of claim 13 , further comprising: obtaining a reference textual dataset pertaining to a textual object of interest and associating with a textual pre-defined quantity target; wherein the reference textual dataset comprises questions/phrases related to an object of interest; and wherein the object of interest comprises masking words representative of the reference textual dataset;

guiding a user to associate variations of the questions/phrases with the reference textual dataset and determining a textual similarity score; wherein the variations of the questions/phrases comprises grammatical variations, active/passive voice variations and/or contextual variations;

classifying in real time, the variations of the questions/phrases as accepted, if the similarity score is below a second pre-defined threshold;

storing the accepted variations of the questions/phrases as training dataset for training a machine learning model; and

repeating the steps of guiding, classifying and storing until the textual pre-defined quantity target criteria is satisfied.

15 . An image curation system for curation of training dataset for training machine learning models comprises:

an image capture device configured to capture images of a scene;

a hardware device comprising one or more storage devices and one or more processing units;

wherein the storage device comprises instructions stored therein that when executed by the one or more processing unit causes the image curation system to:

obtain a reference image frame comprising an object of interest captured using the image capture device;

obtain a reference datapoint for a label or a set of labels from the reference image; wherein the label or set of labels is associated with the object of interest;

annotate the reference datapoint to ascertain an expected annotation for each label or a set of labels; wherein the expected annotation comprises a geometrical shape;

establish the expected annotations and a pre-defined quantity target as seed data for each label or the set of labels;

scale the geometrical shape of the expected annotation based on a predetermined scaling factor;

obtain a position of the scaled geometric shape as an annotation guideline;

shift the position of the annotation guideline by varying the position in a two-dimensional coordinate system and displaying the shifted annotation guideline and a live image on the screen of the image capture device; wherein the live image comprises an image including the object of interest; prompt a user to adjust the live image frame such that the object of interest is tightly encompassed inside the scaled and shifted annotation guideline displayed on the live image screen;

capture a second live image frame and compare the captured second live image frame with the reference image frame to determine a similarity score;

classify, in real time, the captured second live image frame as accepted, if the similarity score is above a pre-defined threshold; store the accepted image frames in the one or more storage device as training dataset for machine learning models;

repeat, for newly captured live image frame, the shift, prompt, capture, classify and store functions until a pre-defined quantity target is satisfied, based on a comparison between the newly captured live image frame and the reference image frame.

16 . The system of claim 15 , wherein comparing the captured second live image frame further comprises comparing the object of interest in the reference frame and captured second live image frame.

17 . The system of claim 15 , wherein the training dataset further comprises audio dataset, image dataset, video dataset and/or textual dataset.

18 . The system of claim 15 , wherein the storage device comprises further instructions causing the image curation system to:

obtain a reference textual dataset pertaining to a textual object of interest and associating with a textual pre-defined quantity target; wherein the reference textual dataset comprises questions/phrases pertaining to an object of interest;

guide a user to associate variations of the questions/phrases with the reference textual dataset and determining a textual similarity score; wherein the variations of the questions/phrases comprises grammatical variations, active/passive voice variations and/or contextual variations;

classify in real time, the variations of the questions/phrases as accepted, if the textual similarity score is below a second pre-defined threshold;

store the accepted variations of the questions/phrases as training dataset for training a machine learning model; and repeat the guide, classify, and store functions until the textual pre-defined quantity target criteria is satisfied.