Method and apparatus for detecting object in image

1 . An object detection method performed by an object detection apparatus, the method comprising:

obtaining, using an object detection model, a result of detecting a target candidate object from an input image;

obtaining, using an error prediction model, a result of detecting an error object from the input image; and

detecting a target object in the input image based on the result of detecting the target candidate object and the result of detecting the error object,

wherein the result of detecting the target candidate object comprises an object score corresponding to a target region of the target candidate object detected in the input image and the target candidate object, and

wherein the result of detecting the error object comprises an error region of the error object detected in the input image, and the error region is spatially defined by a bounding box,

wherein the detecting of the target object comprises:

determining a final score of the target candidate object by applying, to the object score, a weight determined based on an intersection-over-union (IOU) function between the target region of the target candidate object and the error region of the error object; and

detecting the target object in the input image based on the final score,

wherein the object detection method further comprises:

obtaining an additional result of detecting the target candidate object from the input image using an error compensation model,

wherein the additional result comprises an error compensation target region of the target candidate object detected using the error compensation model,

wherein the detecting of the target object comprises:

determining a final score of the target candidate object by applying, to the object score, a first weight determined based on a degree of overlap between the target region of the target candidate object detected using the object detection model and the error region of the error object and a second weight determined based on a degree of overlap between the target region of the target candidate object using the object detection model and the error compensation target region of the target candidate object detected using the error compensation model; and

detecting the target object in the input image based on the final score,

wherein the object detection model does not share a first extraction layers with the error prediction model or the error compensation model, the error prediction model and the error compensation model share a second feature extraction lavers.

2 . The object detection method of claim 1 , wherein the detecting of the target object comprises:

determining the target candidate object as the target object in response to the final score for the target candidate object being equal to or greater than a threshold.

3 . The object detection method of claim 1 , wherein the object detection model provides the result of detecting the target candidate object based on feature data extracted from the input image, and

wherein the error prediction model provides the result of detecting the error object based on another feature data extracted from the input image.

4 . The object detection method of claim 1 , wherein the object detection model and the error prediction model are each based on a separate neural network.

5 . The object detection method of claim 1 , wherein the error prediction model outputs the result of detecting the error object and an additional result of the target candidate object, and

wherein the detecting of the target object comprises detecting the target object in the input image based on the result of detecting the target candidate object, the result of detecting the error object, and the additional result of detecting the target candidate object.

6 . The object detection method of claim 1 , wherein the result of detecting the target candidate object comprises the result of detecting the target candidate object for a plurality of object classes, and

wherein the plurality of object classes comprises a person and a capturing apparatus.

7 . A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method of claim 1 .

8 . An object detection apparatus, comprising:

one or more processors are configured to:

obtain a result of detecting a target candidate object from an input image using an object detection model;

obtain a result of detecting an error object from the input image using an error prediction model; and

detect a target object in the input image based on the result of detecting the target candidate object and the result of detecting the error object,

wherein the result of detecting the target candidate object comprises an object score corresponding to a target region of the target candidate object detected in the input image and the target candidate object,

wherein the result of detecting the error object comprises an error region of the error object detected in the input image, and the error region is spatially defined by a bounding box, and

wherein the one or more processors are further configured to:

determine a final score of the target candidate object by applying, to the object score, a weight determined based on an intersection-over-union (IOU) function between the target region of the target candidate object and the error region of the error object; and

detect the target object in the input image based on the final score,

wherein the object detection method further comprises:

obtaining an additional result of detecting the target candidate object from the input image using an error compensation model,

wherein the additional result comprises an error compensation target region of the target candidate object detected using the error compensation model,

wherein the detecting of the target object comprises:

determining a final score of the target candidate object by applying, to the object score, a first weight determined based on a degree of overlap between the target region of the target candidate object detected using the object detection model and the error region of the error object and a second weight determined based on a degree of overlap between the target region of the target candidate object using the object detection model and the error compensation target region of the target candidate object detected using the error compensation model, and

detecting the target object in the input image based on the final score,

wherein the object detection model does not share a first extraction layers with the error prediction model or the error compensation model, the error prediction model and the error compensation model share a second feature extraction layers.

9 . An image monitoring apparatus, comprising:

one or more processors configured to:

obtain a result of detecting a target candidate object from an image using an object detection model;

obtain a result of detecting an error object from the image using an error prediction model;

detect a target object from the input image based on the result of detecting the target candidate object and the result of detecting the error object; and

execute an operation corresponding to a preset target object in response to the preset target object being detected as the result of detecting the target object,

wherein the result of detecting the target candidate object comprises an object score corresponding to a target region of the target candidate object detected in the input image and the target candidate object, and

wherein the result of detecting the error object comprises an error region of the error object detected in the input image, and the error region is spatially defined by a bounding box, and wherein the one or more processors are further configured to:

determine a final score of the target candidate object by applying, to the object score, a weight determined based on an intersection-over-union (IOU) function between the target region of the target candidate object and the error region of the error object; and

detect the target object in the input image based on the final score,

wherein the object detection method further comprises:

obtaining an additional result of detecting the target candidate object from the input image using an error compensation model,

wherein the additional result comprises an error compensation target region of the target candidate object detected using the error compensation model,

wherein the detecting of the target object comprises:

determining a final score of the target candidate object by applying, to the object score, a first weight determined based on a degree of overlap between the target region of the target candidate object detected using the object detection model and the error region of the error object and a second weight determined based on a degree of overlap between the target region of the target candidate object using the object detection model and the error compensation target region of the target candidate object detected using the error compensation model; and

detecting the target object in the input image based on the final score,

wherein the object detection model does not share a first extraction layers with the error prediction model or the error compensation model, the error prediction model and the error compensation model share a second feature extraction layers.

10 . An object detection apparatus, comprising:

one or more processors are configured to:

extract feature data from an input image using a feature extractor;

detect a first target candidate object from the feature data using an object detection model;

detect an error object from the feature data using an error prediction model;

detect a second target candidate object from the feature data using an error compensation model; and

detect a target object in the input image based on the first target candidate object, the error object, and the second target candidate object,

wherein a result of detecting each target candidate object comprises a first object score corresponding to a target region of the first target candidate object detected in the input image, a second object score corresponding to a target region of the second target candidate object detected in the input image, the first target candidate object and the second target candidate object,

wherein a result of detecting the error object comprises an error region of the error object detected in the input image, and the error region is spatially defined by a bounding box, and

wherein the one or more processors are further configured to:

determine a first final score of the first target candidate object by applying, to the first object score, a weight determined based on an intersection-over-union (IOU) function between a first target region of the first target candidate object and the error region of the error object,

determine a second final score of the second target candidate object by applying, to the second object score, a weight determined based on an intersection-over-union (IOU) function between a second target region of the second target candidate object and the error region of the error object, and

detect the target object in the input image based on the first final score and the second final score,

wherein the object detection method further comprises:

obtaining an additional result of detecting the target candidate object from the input image using an error compensation model,

wherein the additional result comprises an error compensation target region of the target candidate object detected using the error compensation model,

wherein the detecting of the target object comprises:

determining a final score of the target candidate object by applying, to the object score, a first weight determined based on a degree of overlap between the target region of the target candidate object detected using the object detection model and the error region of the error object and a second weight determined based on a degree of overlap between the target region of the target candidate object using the object detection model and the error compensation target region of the target candidate object detected using the error compensation model; and

detecting the target object in the input image based on the final score,

wherein the object detection model does not share a first extraction layers with the error prediction model or the error compensation model, the error prediction model and the error compensation model share a second feature extraction layers.

11 . The object detection apparatus of claim 10 , further comprising a memory configured to store instructions;

wherein the one or more processors are further configured to execute the instructions to configure the one or more processors to:

receive the input image;

extract the feature data from the input image using the feature extractor;

detect the first target candidate object from the feature data using the object detection model;

detect the error object from the feature data using the error prediction model;

detect the second target candidate object from the feature data using the error compensation model; and

detect the target object in the input image based on the first target candidate object, the error object, and the second target candidate object.

12 . The object detection apparatus of claim 10 , wherein the feature extractor comprises a plurality of feature extractors,

wherein a first feature extractor of the plurality of feature extractors provides the feature data of the first target candidate object,

wherein a second feature extractor of the plurality of feature extractors provides the feature data of the error object,

wherein a third feature extractor of the feature extractors provides the feature data of the second target candidate object.

13 . The object detection method of claim 1 , wherein the error prediction model is trained to detect an error caused by the object detection model by analyzing the result of detecting the target candidate object and configured to supplement the error by modifying the result of detecting the target candidate object.

14 . The object detection method of claim 1 , wherein the error prediction model is trained to detect an error caused by the object detection model by analyzing the result of detecting the target candidate object and configured to supplement the error by adjusting the final score of the target candidate object.