Methods and systems for the automated quality assurance of annotated images
A framework in which annotated images can be analyzed in small batches to learn and distinguish between higher-quality annotations and lower-quality annotations, especially in the case of manual annotations for which quality assurance is desired. This framework is extremely generalizable and can be used for indoor images, outdoor images, medical images, etc., without limitation. An echo state network (ESN) is provided as a special case of semantic segmentation model that can be trained using as few as tens of annotated images to predict semantic regions and provide metrics that can be used to distinguish between higher-quality annotations and lower-quality annotations.
1. A method, comprising:
training a plurality of parallel semantic segmentation models on an initial annotated dataset;
generating a plurality of annotation regional proposals for a batch of images;
computing a confidence metric indicating a degree of agreement between the plurality of annotation regional proposals;
providing the batch of images to a first manual annotator and a second manual annotator to generate a first manual annotation set and a second manual annotation set and determining a first confidence score associated with the first manual annotator related to the first manual annotation set and a second confidence score associated with the second manual annotator related to the second manual annotation set; and
assessing a preferred of the first manual annotator and the second manual annotator by comparing the first confidence score and the second confidence score;
wherein the confidence metric is computed with the plurality of annotation regional proposals as inputs by computing intersection-over-union (IOU) and Dice (F1) scores for regional proposal pairs and computing a confidence (confid_p) comprising a mean over a variance, where a denominator is a standard deviation between paired IOU or F1 scores.
2. The method of claim 1 , wherein the plurality of parallel semantic segmentation models comprises a plurality of parallel echo state network models.
3. The method of claim 1 , wherein the initial annotated dataset comprises fewer than 100 annotated images.
4. The method of claim 1 , wherein the first confidence score and the second confidence score are each determined by finding pairs of IOUs and F1s for each manual annotated image of each manual annotation set.
5. The method of claim 1 , wherein assessing the preferred of the first manual annotator and the second manual annotator comprises:
determining whether the first confidence score and the second confidence score are below a predetermined threshold and declaring a quality assessment automation failure and providing the first manual annotation set and the second manual annotation set to a master manual annotator for analysis if determined that the first confidence score and the second confidence score are below the predetermined threshold; and
determining whether the first confidence score and the second confidence score are different to a predetermined degree and declaring a quality assessment automation success and selecting the preferred of the first manual annotator and the second manual annotator based on the higher of the first confidence score and the second confidence score if determined that the first confidence score and the second confidence score are different to the predetermined degree.
6. The method of claim 1 , wherein the method is used in training an autonomous driving/advanced driver assistance system of a vehicle.
7. The method of claim 1 , wherein the method is executed for one of indoor images, outdoor images, and medical images.
8. A non-transitory computer readable medium stored in a memory and executed by a processor to execute the steps, comprising:
training a plurality of parallel semantic segmentation models on an initial annotated dataset;
using the trained plurality of parallel semantic segmentation models, generating a plurality of annotation regional proposals for a batch of images;
using the plurality of annotation regional proposals, computing a confidence metric indicating a degree of agreement between the plurality of annotation regional proposals;
providing the batch of images to a first manual annotator and a second manual annotator to generate a first manual annotation set and a second manual annotation set and determining a first confidence score associated with the first manual annotator related to the first manual annotation set and a second confidence score associated with the second manual annotator related to the second manual annotation set; and
assessing a preferred of the first manual annotator and the second manual annotator by comparing the first confidence score and the second confidence score;
wherein the confidence metric is computed with the plurality of annotation regional proposals as inputs by computing intersection-over-union (IOU) and Dice (F1) scores for regional proposal pairs and computing a confidence (confid_p) comprising a mean over a variance, where a denominator is a standard deviation between paired IOU or F1 scores.
9. The non-transitory computer readable medium of claim 8 , wherein the plurality of parallel semantic segmentation models comprises a plurality of parallel echo state network models.
10. The non-transitory computer readable medium of claim 8 , wherein the initial annotated dataset comprises fewer than 100 annotated images.
11. The non-transitory computer readable medium of claim 8 , wherein the first confidence score and the second confidence score are each determined by finding pairs of IOUs and F1s for each manual annotated image of each manual annotation set.
12. The non-transitory computer readable medium of claim 8 , wherein assessing the preferred of the first manual annotator and the second manual annotator comprises:
determining whether the first confidence score and the second confidence score are below a predetermined threshold and declaring a quality assessment automation failure and providing the first manual annotation set and the second manual annotation set to a master manual annotator for analysis if determined that the first confidence score and the second confidence score are below the predetermined threshold; and
determining whether the first confidence score and the second confidence score are different to a predetermined degree and declaring a quality assessment automation success and selecting the preferred of the first manual annotator and the second manual annotator based on the higher of the first confidence score and the second confidence score if determined that the first confidence score and the second confidence score are different to the predetermined degree.
13. The non-transitory computer readable medium of claim 8 , wherein the steps are used in training an autonomous driving/advanced driver assistance system of a vehicle.
14. The non-transitory computer readable medium of claim 8 , wherein the steps are executed for one of indoor images, outdoor images, and medical images.
15. A system, comprising:
a processor executing an algorithm stored in a memory operable for:
training a plurality of parallel semantic segmentation models on an initial annotated dataset;
using the trained plurality of parallel semantic segmentation models, generating a plurality of annotation regional proposals for a batch of images;
using the plurality of annotation regional proposals, computing a confidence metric indicating a degree of agreement between the plurality of annotation regional proposals;
providing the batch of images to a first manual annotator and a second manual annotator to generate a first manual annotation set and a second manual annotation set and determining a first confidence score associated with the first manual annotator related to the first manual annotation set and a second confidence score associated with the second manual annotator related to the second manual annotation set; and
assessing a preferred of the first manual annotator and the second manual annotator by comparing the first confidence score and the second confidence score;
wherein the confidence metric is computed with the plurality of annotation regional proposals as inputs by computing intersection-over-union (IOU) and Dice (F1) scores for regional proposal pairs and computing a confidence (confid_p) comprising a mean over a variance, where a denominator is a standard deviation between paired IOU or F1 scores.
16. The system of claim 15 , wherein the plurality of parallel semantic segmentation models comprises a plurality of parallel echo state network models.
17. The system of claim 15 , wherein the initial annotated dataset comprises fewer than 100 annotated images.
18. The system of claim 15 , wherein the first confidence score and the second confidence score are each determined by finding pairs of IOUs and F1s for each manual annotated image of each manual annotation set.
19. The system of claim 15 , wherein assessing the preferred of the first manual annotator and the second manual annotator comprises:
determining whether the first confidence score and the second confidence score are below a predetermined threshold and declaring a quality assessment automation failure and providing the first manual annotation set and the second manual annotation set to a master manual annotator for analysis if determined that the first confidence score and the second confidence score are below the predetermined threshold; and
determining whether the first confidence score and the second confidence score are different to a predetermined degree and declaring a quality assessment automation success and selecting the preferred of the first manual annotator and the second manual annotator based on the higher of the first confidence score and the second confidence score if determined that the first confidence score and the second confidence score are different to the predetermined degree.
20. The system of claim 15 , wherein the processor executes the algorithm for one of indoor images, outdoor images, and medical images.