Method for determining eye measurements

Abstract

The invention relates to a method (1) for determining eye measurements comprising the following steps: - selecting (100) an image in a video stream, - identifying (200) a known object in the image, - detecting (300) a face in the image, said face comprising two sides, each side comprising an eye and an ear, - determining (400) an eye measurement of the face using a reference length of the known object, characterized in that the step of selecting the image comprises the following sub-steps, - detecting (101) the face in the video stream, - determining (106) for each side of the face an eye y-coordinate, and an ear y-coordinate, - determining (107) for each side of the face an eye-ear length, - selecting (110) the image in the video stream so that the eye y-coordinates, the ear y-coordinates and the eye-ear lengths meet certain criteria.

Description

Description Title of the invention: Method for determining ocular measurements Technical field of the invention

[0001] The field of the invention is that of taking eye measurements in the context of the manufacture of corrective lenses and the mounting of corrective lenses in eyeglass frames. Prior art

[0002] When manufacturing corrective lenses and fitting corrective lenses into eyeglass frames, it is necessary to know ocular measurements such as pupillary distance, monopupillary distance and heights as accurately as possible.

[0003] Pupillary distance, referred to as PD in this text, is the distance between the centers of the pupils when a subject is looking to infinity. A subject is a wearer or future wearer of glasses. A subject is considered to be looking to infinity when the subject is looking at an object far enough away from them that the direction of gaze of one of their eyes is approximately parallel to the direction of gaze of their other eye.

[0004] The monopupillary distance, called mono PD in this text, is the distance between the projection of the center of one of the pupils in the plane of the frame and the center of the frame.

[0005] Heights are the distances between the projection of the center of the pupils in the plane of the frame and the bottom of the frame. The bottom of the frame means the inside of the internal dragee.

[0006] Eye measurements are traditionally determined by an optical professional in the presence of a wearer or future wearer of glasses. The optical professional generally traces, using a felt-tip pen, the pupil center directly on the lens while facing the wearer of glasses. This method is subject to human error and suffers from a lack of precision. In order to increase the accuracy of determining eye measurements and to avoid the wearer of glasses and / or the optical professional having to move, techniques for determining eye measurements using a consumer sensor are described in particular in patent FR30088005. These techniques are based on a transfer of 2D measurements from a face and a known object present in an image acquired by the consumer sensor.

[0007] The quality of eye measurements determined using these techniques is, however, degraded by the orientation of the user's face and the known object in the image(s) used. Additionally, when the focal length of the sensor used is not unknown, the position of the glasses wearer and the known object in front of the sensor also degrades the quality of the determined eye measurements. Presentation of the invention

[0008] The present invention overcomes the aforementioned drawbacks by providing a method for determining ocular measurements.

[0009] Said method for determining ocular measurements comprises the following steps: - select an image in a video stream, - identify a known object in the selected image, - detect a face in the selected image, said face having two sides, called left side and right side, each side having an eye and an ear, - determine an eye measurement of the face using a reference length of the known object, measured in the selected image.

[0010] According to the invention, the image selection step comprises the following sub-steps, - detect face in video stream, - determine for each side of the face an ordinate of a first reference point of the eye, called the eye ordinate, and an ordinate of a first reference point of the ear, called the ear ordinate, - determine for each side of the face a length between a second reference point of the eye and a second reference point of the ear, called eye-ear length, - select the image in the video stream so that for each side of the face, a difference between the eye ordinate and the ear ordinate is less than an ordinate threshold and so that a difference between the eye-ear length on one side of the face and the eye-ear length on the other side of the face is less than a length threshold.

[0011] The reference length of the known object is understood to be a known length which, once measured in the image, will allow, by length transfer, to establish an ocular measurement such as, for example, the PD.

[0012] A reference point is a fixed point on a part of the face.

[0013] The ordinate of a reference point is the distance, in an image, between the reference point and the lower boundary of the image. In the context of a rectangular image, for example, by placing an orthogonal coordinate system in the lower left corner of the image, this distance is the ordinate of the reference point in the orthogonal coordinate system.

[0014] The first reference point and the second reference point of the eye can be confused and therefore form only one point.

[0015] The first reference point and the second reference point of the ear can be confused and therefore form only one point.

[0016] The ear on each side of the face is considered to be towards the outside of the face relative to the eye on the corresponding side of the face.

[0017] By using a low length threshold, generally close to zero, the selected image advantageously presents the face with a lateral orientation facing the sensor, that is to say, the face of the glasses wearer is neither turned to the left nor turned to the right. In this way, the glasses wearer will be able to place the known object resting on his face and a length transfer can be made between the reference length of the known object and for example the distance between the pupils of the glasses wearer, the PD, with a minimum of error in a horizontal direction.

[0018] By using a low ordinate threshold, the selected image presents the face with a vertical orientation facing the sensor, that is, the face of the eyeglass wearer is neither turned upwards nor turned downwards. In this way, the eyeglass wearer will be able to place the known object on his face and a length transfer can be made between the reference length of the known object and, for example, the distance between the pupils of the eyeglass wearer, the PD, with a minimum of error in a vertical direction.

[0019] The selected image meeting both criteria, in lateral orientation and in vertical orientation, allows a more precise ocular measurement than with prior art techniques.

[0020] Such steps therefore allow the method according to the invention to select an image of the face of a glasses wearer or future glasses wearer that is optimal for determining eye measurement by 2D measurement transfer, then to determine an eye measurement with high precision.

[0021] In particular embodiments, the invention may further comprise one or more of the following characteristics, taken individually or in any technically possible combination.

[0022] According to one embodiment, for each side of the face, the first point and the second reference point of the eye are the outer corner of the corresponding eye and the first point and the second reference point of the ear are the point of the corresponding ear on which a branch of a pair of glasses would rest.

[0023] The outer corner of the eye and the point on the ear on which a branch of a pair of glasses would rest are points on the eye and ear that can be identified with high precision by image processing.

[0024] Additionally, when the outer corner of the eye is the first and second reference point of the eye and when the point of the ear on which a branch of a pair of glasses would rest is the first and second reference point of the eye, ear reference, low length and ordinate thresholds allow the selected image to present the face in an optimal orientation.

[0025] According to one embodiment, the ordinate threshold and the length threshold are less than 2% of a face width.

[0026] A face width is defined as the length corresponding to the width of the face as observed in an image of the video stream, for example, the distance between a left end and a right end of the face. An ordinate threshold and a length threshold less than 2% of a face width, for example, for an image produced by a sensor with a resolution equal to or equivalent to 1280x720, ensure a centered Gaussian distribution of the error on the eye measurement, a standard deviation of the error on the eye measurement of 0.8 mm and a maximum error of 2 mm. An image with a resolution higher than 1280x720 will further reduce the error on the eye measurement.

[0027] According to one embodiment, the image selection in the video stream is performed by minimizing for each side the difference between the eye ordinate and the ear ordinate and the difference between the eye-ear length on one side of the face and the eye-ear length on the other side of the face.

[0028] According to one embodiment, the image selection step comprises the following sub-step: - displaying the video stream with at least one visual indicator, each at least one visual indicator being a function of at least one of the following: the right eye ordinate, the left eye ordinate, the right ear ordinate, the left ear ordinate, the right eye-ear length, and the left eye-ear length.

[0029] The glasses wearer or future glasses wearer will be able to watch the video stream and improve their orientation in front of the sensor using at least one visual indicator. The above sub-step therefore makes it possible to reduce the measurement time by guiding the glasses wearer to quickly orient themselves in front of the sensor in order to produce an image suitable for taking measurements.

[0030] According to one embodiment, the sub-step of detecting the face in the video stream comprises the following sub-steps: - estimate a 3D face model corresponding to the face, - match a position and orientation of the 3D face model with the face in the video stream.

[0031] Such additional sub-steps improve the accuracy of positioning the eye and ear reference points. Indeed, by using the 3D face model, estimated to match the face in the video stream and matching the position of the 3D face model orientation with the face, The reference points are positioned on the face with greater accuracy. Increasing the positioning accuracy of the reference points on the face further improves the orientation of the face in the selected image and thus further reduces the error in eye measurements.

[0032] According to one embodiment, the sub-step of detecting the face in the video stream comprises the following sub-step: - refining the estimation of the 3D face model from several refining images of the video stream, said refining images having different orientations of the 3D face model.

[0033] Such a sub-step of refining the estimation of the 3D model makes it possible to improve the quality of the estimation of the 3D face model in an iterative manner and thus to further improve the positioning accuracy of the reference points of the eyes and ears.

[0034] According to one embodiment, the image selection step comprises the following sub-steps: - determine a face size, - determine a center of the face,

[0035] and the image is selected from the video stream such that the size of the face is less than a size threshold and such that a distance between the center of the face and a center of the image is less than a centering threshold.

[0036] The distortions caused by a sensor focal length are limited to the center of an image from the sensor and increase as they move away from the center of the image. In the case of a sensor whose distortions of the captured images caused by the sensor focal length are not known, that is to say in the case of an uncalibrated sensor, a low size threshold and a low centering threshold make it possible to limit the errors in determining ocular measurement, caused by the focal length, from the selected image.

[0037] In this text, focal length is understood to mean the focal length.

[0038] According to one embodiment, the size threshold is less than 20% of an image size and the centering threshold is less than 2% of the face width.

[0039] In the case of an uncalibrated sensor, with a resolution equal to or equivalent to 1280x720, a size threshold lower than 20% of an image size, for example when the face size is a diagonal length of the face and the image size a diagonal length of the image, and a centering threshold lower than 2% of the face width, make it possible to guarantee a controlled error on the ocular measurement. Controlled error means an error respecting a centered Gaussian distribution with a standard deviation of 0.8 mm and a maximum of 2 mm. An image with a resolution higher than 1280x720 will further reduce the error on the ocular measurement. Advantageously, when the size threshold is between 20% and 10% of the image size, for a consumer sensor, a distance between the user and the consumer sensor is between approximately 60 cm and approximately 90 cm, which corresponds to an outstretched arm distance. The outstretched arm distance between the user and the sensor makes it possible, for example, to determine the PD measurement, thanks to the addition of a statistical bias. The statistical bias corresponds to the difference between a pupillary distance for a direction of gaze towards an object at the outstretched arm distance and a pupillary distance for a direction of gaze towards infinity. The statistical bias could for example be learned from a large number of users of the protocol according to the invention and whose PD will have also been measured when the direction of their gaze is towards infinity.

[0040] According to one embodiment, the image selection step comprises the following sub-step: - displaying the video stream with at least one visual indicator, each at least one visual indicator being a function of at least one of the following: face size, face center, right eye ordinate, left eye ordinate, right ear ordinate, left ear ordinate, right eye-ear length, and left eye-ear length.

[0041] The glasses wearer or future glasses wearer will be able to watch the video stream and improve their orientation and positioning in front of the sensor using at least one visual indicator. The above sub-step therefore makes it possible, in the case of an uncalibrated sensor, to reduce the measurement time by guiding the glasses wearer to quickly orient and position themselves in front of the sensor so as to produce an image suitable for taking measurements.

[0042] According to one embodiment, the known object is a pair of glasses worn on the face by a user whose face is detected in the image and the video stream.

[0043] Using a statistical model of glasses constructed from the 3D scan of many pairs of glasses, the type and shape of the glasses can be recognized in the selected image, for example, a pair type with studs or a pair type with large lenses. Depending on the type and shape, the dimensions of the pair of glasses worn by the user, also called the glasses wearer, can be estimated, for example, using an optimization algorithm to adjust a projection of the statistical model of the pair of glasses onto a segmentation mask of the pair of glasses. The segmentation mask can be obtained using an automatic deep learning algorithm. In this way, the reference length can be obtained from the pair of glasses.

[0044] According to one embodiment, the pair of glasses has been previously scanned.

[0045] In the case of a pair of glasses previously scanned, the pair of glasses present in the image is known and its dimensions are available in a database constructed from 3D scans of pairs of glasses. The reference length can be obtained from the database.

[0046] According to one embodiment, a glass width and a bridge width of the pair of glasses are previously known.

[0047] In the case of a pair of glasses not previously scanned but whose lens width and bridge width are known, these dimensions are commonly called by the English term "frame marking", and are often written inside a branch of the pair of glasses. The glasses wearer will be able, using an interface, to indicate the frame marking values ​​and these values, for example, corrected for the deformation caused by the curve of the pair of glasses, can be used as a reference length to determine the ocular measurement. Presentation of figures

[0048] The invention will be better understood by reading the following description, given as a non-limiting example, and with reference to the figures:

[0049] [Fig.l] a schematic representation of an example of a method according to the invention,

[0050] [Fig.2] a representation of images in the video stream,

[0051] [Fig.3] another representation of images in the video stream,

[0052] [Fig.4] a representation of an image containing a known object,

[0053] [Fig.5] a schematic representation of an example of ocular measurement,

[0054] [Fig.6] a representation of a scanned scope correspondence with a scope in an image,

[0055] [Fig.7] a schematic representation of another example of ocular measurement,

[0056] In these figures, identical references from one figure to another designate identical or similar elements. For reasons of clarity, the elements represented are not necessarily to the same scale, unless otherwise indicated.

[0057] Detailed description of particular embodiments of the invention

[0058] [Fig.l] is a schematic representation of a non-limiting example of implementation of a method 1 according to the invention. The method 1 comprises a step 100 of selecting an image 2 in a video stream. Once image 2 is selected, a known object 3 is identified in image 2 during a step 200, and a face 4, for which an eye measurement is to be obtained, is detected in image 2 during a step 300. Then an eye measurement such as the PD can be determined during a step 400 in the following manner: a reference length of the known object is found in image 2, the distance to be measured is found in image 2, for example the distance between two pupils of the face (PD), then the distance between the two pupils is calculated using, for example, a rule of three, from the distance to be measured in the image, the reference length found in image 2 and the known actual value of the reference length.

[0059] The step 100 of selecting the image 2 in the video stream comprises the following sub-steps. A sub-step 101 of detecting the face 4 in the video stream is visible in [Fig.1]. This step may be carried out, for example, using a facial recognition model. The face 4 has two sides, a left side and a right side, each side of the face comprising an eye 5 and an ear 6. Reference points may be identified for each eye and for each ear. A first reference point 50 of the eye 5 may be, for example, the center of the pupil and a second reference point 51 of the eye may be, for example, the inner corner of the eye 5. A first reference point 60 of the ear 6 may be, for example, the top of the auricle and a second reference point 61 of the ear 61 may be, for example, the entrance to the auditory canal.The location of the reference points on the face 4 is determined, for example, using a facial recognition model.

[0060] A determination sub-step 106 is then carried out. In this step, for each side of the face 4, an ordinate of the first reference point 50 of the eye 50, called the eye ordinate, and an ordinate of the first reference point 60 of the ear 6, called the ear ordinate, are determined.

[0061] The ordinate of a reference point is the distance, in an image, between the reference point and, for example, the lower limit of the image. In the context of a rectangular image, by placing an orthogonal coordinate system in the lower left corner of the image, this distance is the ordinate of the reference point in the orthogonal coordinate system.

[0062] A determination sub-step 107 is then carried out. In this step, for each side of the face 4, a length between the second reference point 51 of the eye 5 and the second reference point 61 of the ear 6, called the eye-ear length, is determined.

[0063] The length between two reference points of the face 4 is, for example, a number of pixels in the image 2 between the two reference points.

[0064] Then, image 2 is selected during a sub-step 110 in the video stream so that image 2 meets the following criteria: - for each side of the face 4, a difference between the eye ordinate and the ear ordinate is less than an ordinate threshold, - a difference between the eye-ear length on one side and the eye-ear length on the other side is less than a length threshold.

[0065] In this way, by setting a low ordinate threshold, face 4 in image 2 is oriented vertically facing a sensor producing the video stream. By "vertically oriented facing the sensor" we mean that face 4 is neither facing up nor down in image 2.

[0066] In this way also, by defining a low length threshold, face 4 in image 2 is oriented laterally facing the sensor. By "oriented laterally facing the sensor" we mean that face 4 is neither oriented to the left nor to the right in image 2.

[0067] Alternatively, image 2 is selected by minimizing the difference between the eye ordinate and the ear ordinate for each side and minimizing the difference between the eye-ear length on one side and the eye-ear length on the other side.

[0068] Alternatively, other facial reference points may be used to carry out the method according to the invention. Any reference points highlighting facial symmetry may be used, for example, a nose reference point and a jaw or mouth reference point for each side of the face.

[0069] A user, whose face 4 is present in image 2, and for whom an eye measurement is to be determined, can thus place the known object 3 on his face so that the reference length of the object is in a plane parallel to a plane of his face. Image 2 thus selected will make it possible to carry out the determination 400 of eye measurement with greater precision than that obtained using prior art techniques.

[0070] An additional sub-step of verifying the immobility of the user in the video stream can be added in the step of selecting image 2. This will avoid the selection of an image 2 in which face 4 would be blurred.

[0071] Another additional sub-step could be a verification of image quality according to sharpness and contrast due, in particular, to brightness or sensor quality. This would avoid selecting an image 2 in which face 4 would not be clear.

[0072] A case of absence of visibility of an ear can also be covered by adding a condition to the selection of image 2. This condition could be, for example, that for each side of the face the ear is positioned relative to the eye towards the outside of the corresponding side of the face. An ear reference point detector can also provide visibility information for each reference point that can be used to perform face registration.

[0073] According to one embodiment, the first reference point 50 of the eye 5 and the second reference point 51 of the eye 5 are merged and are the outer corner of the corresponding eye 5 and the first reference point 60 of the ear 6 and the second point 61 of the ear 6 are merged and are the point of the corresponding ear 6 on which a branch of a pair of glasses would rest. The outer corner of the eye and the point of the ear on which a branch of a pair of glasses would rest are points of the eye and the ear identifiable by image processing, using for example a facial recognition model, with good precision. In other words, the location of these reference points on face 4 in the video stream can be precisely identified. In addition, these reference points are particularly suitable for allowing the selection of image 2 which will present face 4 in an optimal orientation.

[0074] According to one embodiment, the ordinate threshold and the length threshold are less than 2% of a face width. In this way, for a resolution sensor equal to or equivalent to 1280x720, the orientation of the face 4 in the image 2 allows an error in determining 400 of ocular measurement respecting a centered Gaussian distribution, with a standard deviation of 0.8 mm and a maximum of 2 mm. According to one embodiment, the ordinate threshold and the length threshold are less than 1% of the face width. In this way, the standard deviation and the maximum of the error in determining 400 of ocular measurement are further reduced.

[0075] According to one embodiment, the step 100 of selecting the image 2 comprises a sub-step 108 of displaying at least one visual indicator 7 in the video stream. The at least one visual indicator 7 is a function of at least one of the following elements: - the right eye ordinate, - the left eye ordinate, - the right ear ordinate, - the left ear ordinate, - the length of the right eye and ear, - the length from eye to left ear.

[0076] The at least one visual indicator 7 may be, for example, a vertical segment, displayed in the video stream at the eye 5 on one side of the face 4, the length of which depends on the difference between the eye ordinate and the ear ordinate for the corresponding side of the face. The at least one visual indicator 7 may also be, for example, a horizontal segment displayed in the video stream at an ear 6 of the face 4 and the length of which depends on the difference between the eye-ear length on one side of the face and the eye-ear length on the other side of the face.

[0077] This sub-step 108 of displaying at least one visual indicator 7 makes it possible to guide the user so that he or she orients his or her face optimally in front of the video stream sensor. In the absence of a visual indicator 7, image 2 can be selected according to the relevant criteria. The method according to the invention will, however, last longer, which will occupy the user for longer, the latter taking more time to orient his or her face appropriately.

[0078] According to one embodiment, the sub-step 101 of detecting the face 4 in the video stream comprises a sub-step 102 of estimating a 3D face model corresponding to the face 4 and a sub-step 103 of matching a position and an orientation of the 3D face model with the face 4 in the video stream. The use of the 3D face model for the detection 101 of the face 4 in the video stream allows a rapid convergence of the detection 101 of the face 4 in the video stream, i.e. the face detection time is reduced compared to other facial recognition methods. Additionally, the positioning of the eye and ear reference points on the face 4 is more accurate thanks to the use of the 3D face model. Said 3D face model may be, for example, a model constructed by machine learning, consisting of a 3D face model skeleton. The dimensions of the 3D face model skeleton will be dynamically adjusted using one or more images of the face 4. The correspondence of the position and orientation of the 3D face model with the face 4 in the video stream is also called face alignment in this text.

[0079] Advantageously, when an ear 6 of the user is masked, for example by hair or a pair of glasses, the reference point(s) of the masked ear can still be positioned using the 3D face model and face alignment.

[0080] According to one embodiment, the sub-step 101 of detecting the face 4 in the video stream comprises an additional sub-step 109 of refining the estimation of the 3D face model. To refine the estimation of the 3D face model, several refining images will be used. The refining images present the face 4 in different orientations. Such a refining step 109 makes it possible to improve the estimation of the 3D face model, which makes it possible to further improve the accuracy of the positioning of the reference points on the face 4 and thus to allow the selection 110 of an image 2, further reducing the error 400 in determining the ocular measurement.

[0081] According to one embodiment, the step 100 of selecting the image 2 comprises a sub-step 104 of determining a face size and a sub-step 105 of determining a face center. In this embodiment, the sub-step 110 of selecting the image 2 in the video stream comprises the following additional criteria: - the face size is below a size threshold, - the distance between the center of the face and a center of image 2 is less than a centering threshold.

[0082] When the distortion caused by the focal length of the sensor is unknown, that is, when the sensor is uncalibrated, which is the case for consumer sensors, a low centering threshold and a low size threshold make it possible to obtain a position of face 4 in image 2 and a depth between the user and the sensor which very strongly limit the distortion of face 4 in image 2. In this way, the selected image 2 will make it possible to determine an ocular measurement with very little error. caused by the focal length of the sensor.

[0083] When the deformation caused by the focal length of the sensor is known, the error caused by this deformation on the determination of the eye measurement can be corrected; the centering and size of face 4 in image 2 will, in this case, have very little influence on the error in determining the eye measurement.

[0084] According to one embodiment, the size threshold is less than 20% of a size of the image 2 and the centering threshold is less than 2% of the width of the face. In this way, for an uncalibrated sensor with a resolution equal to or equivalent to 1280x720, the centering and the size of the face 4 in the image 2 allow an error in determining 400 of the eye measurement respecting a Gaussian distribution. The Gaussian distribution of the error is centered, its standard deviation is 0.8 mm and its maximum is 2 mm. According to one embodiment, the size threshold is between 20% of the size of the image 2 and 10% of the size of the image 2 and the centering threshold is less than 1% of the width of the face. In this way, the standard deviation and the maximum of the error in determining 400 of the eye measurement are further reduced. The image size could be, for example, the horizontal dimension of the image.In this case, the face size 4 will be the horizontal dimension of the face or face width 4.

[0085] According to one embodiment, the at least one visual indicator 7 is a function of at least one of the following additional elements: - the size of the face, - centering the face.

[0086] The at least one visual indicator 7 may be, for example, a frame around the face 4 whose color varies according to the size of the face and the centering of the face; this color variation may be a shade between red and green. The red color indicates, for example, a size of the face greater than the size threshold and a distance between the center of the face and the center of the image greater than the centering threshold. The green color indicates, for example, a size of the face less than the size threshold and a distance between the center of the face and the center of the image less than the centering threshold. Such a visual indicator 7, displayed in the video stream, guides the user so that the latter positions his face with a depth and centering facing the sensor adapted to obtaining an image 2 allowing the determination 400 of ocular measurement with good precision despite the fact that the deformation caused by the focal length of the sensor is not known.The visual indicator 7 thus makes it possible to reduce the time required to carry out the method according to the invention in the case of an uncalibrated sensor, the user being able to position his face more quickly in an appropriate manner.

[0087] According to one embodiment, the known object 3 identified in image 2 during step object identification 200 is a pair of glasses 31 worn on the face by the user for which an eye measurement is to be determined. If the pair of glasses 31 is not known, a statistical model may for example be used to estimate the actual dimensions of the pair of glasses and thus dispose of the reference length used for the determination 400 of the eye measurement. The statistical model is, for example, a model constructed from the 3D scan of numerous pairs of glasses which, for a type and shape of pair of glasses, makes it possible to estimate the dimensions of the pair of glasses from the dimensions of scanned pairs of glasses of the same type and similar shape.

[0088] According to one embodiment, the pair of glasses identified in image 2 has been previously scanned. This is a known pair of glasses whose dimensions have been measured during a scan. A 3D model of glasses constructed during the scanning of the pair of glasses can be matched 207 with the pair of glasses 31 in image 2 and the reference length can thus be determined from the pair of glasses 31 with additional precision than in the previous embodiment, the one in which the pair of glasses is not known.

[0089] According to one embodiment, a lens width and a bridge width of the pair of glasses identified in image 2 are known. If the pair of glasses 31 is not known but the dimensions, also called "frame marking" are available, the user can indicate the lens width and the bridge width using an interface. A step 205 of correcting the curve of the pair of glasses 31 in image 2 can then be carried out in order to obtain the reference length with good precision. Curve is understood to mean the curvature of the face of the pair of glasses generally created when adjusting the pair of glasses to its wearer. Frame marking reference points of the pair of glasses 31 can be identified in image 2 using a detector resulting from deep learning.The frame marking reference point of the pair of glasses is understood to mean points of the pair of glasses 31 in the image allowing the lens width and the bridge width to be identified. Alternatively, a model of the pair of glasses allows the frame marking reference points to be identified more precisely by integrating parameters linked to depths in a scene corresponding to image 2. The scene could for example be made up of several images. Without correction of the curve of the pair of glasses, the reference length can still be obtained from the frame marking but with less precision.

[0090] According to one embodiment, the known object 3 is not a pair of glasses. The object 3 could for example be a standard-sized card, such as a credit card, a loyalty card or a biometric identity card. The known object 3 could also be any other object whose dimensions are previously known and which could be positioned in relation to face 4 in a known manner.

[0091] The step 200 of identifying the known object 3 may, for example, include a step 201 of testing for the presence of a standard card 30 in the image 2. In the absence of a card 30, a step 202 of testing for the presence of a pair of glasses 31 in the image 2 may be carried out, then, if a pair of glasses 31 is identified, a step 203 will determine whether the pair of glasses 31 is known. Step 203 may include the use of an algorithm derived from machine learning. This algorithm uses a segmentation mask of the pair of glasses. The segmentation mask may be obtained by a deep learning method. Then, the 3D model of the pair of glasses is optimized such that a projection of the 3D model of the pair of glasses best fits the segmentation mask. This optimization is also called glasses alignment.Once the eyeglass alignment has been carried out, a measurement report of the dimensions of the eyeglass projected onto the image and of the distance between pupil reference points can be carried out, using expected biases. Pupil reference points are understood to mean points of the image positioned at the center of the pupils. The expected biases are, for example, a vision bias at 80 cm, a lens-eye distance bias and an estimated curve bias. If the pair of eyeglasses 31 is not known, a question 204 may be asked to the user asking him if he has frame marking and then inviting him, if he responds positively, to indicate the frame marking values ​​in an interface. In the absence of frame marking, a step 206 of using the statistical model may be carried out.

[0092] Alternatively, the user is known and the frame marking of the frame he is wearing in the image is already entered, for example thanks to a previous order for a pair of glasses from the user.

[0093] [Fig. 2] is a representation of images in the video stream. A user's face 4 is present in image 2. The ear reference points 60 and 61 are merged in this example. The eye reference points 50 and 51 are also merged in this example. Visual indicators 7 are visible in the images, they guide the user so that the user can optimally orient his face in front of the sensor.

[0094] [Fig. 3] is another representation of images in the video stream. A user's face 4 is visible. A visual indicator 7 is also visible. It guides the user so that the user can optimally position their face in front of the sensor. That is, so that they can center and distance their face sufficiently from the sensor to limit distortions caused by the sensor's focal length.

[0095] [Fig.4] is a representation of image 2 comprising a known object 3. A standard card 30 is visible in image 2, the user positions the card 30 adequately on his face. By "adequately" is meant that a face plane of the card 30 is parallel to the user's face 4.

[0096] [Fig.5] is a schematic representation of an example of ocular measurement. In the absence of a pair of glasses 31 in image 2, the mono pupillary distances 80, mono PD, can be determined using a main bridge of the nose. The PD can be estimated without identifying any other elements of the face 4 than the pupils of the eyes 5.

[0097] [Fig.6] is a representation of a correspondence 207 of a 3D model of a pair of glasses with a pair of glasses in an image, also called glasses alignment. The 3D model of glasses is either derived from a previously scanned known pair of glasses, or from the statistical model of the pair of glasses, or from a model constructed using frame marking. A 2D projection of the 3D model of the pair of glasses could, for example, be produced and this projection could be positioned so as to correspond with the pair of glasses 31.

[0098] A linear parametric model of a pair of glasses can be used to perform this step. This model can be described as a linear combination of a mean and modes.

[0099] [Math.l]

[0100] Considering a following projection function,

[0101] [Math.2] Prof f ( p3D ) ~ K { [ RT ] p3D

[0102] a function to minimize to achieve the correspondence can be written,

[0104] with the following condition: Schematic representation of another example of ocular measurement. In this example, the known object 3 is a pair of glasses 31. The mono PD 80 can be determined using the center of the pair of glasses (the center of the bridge) with better precision than using the main ridge of the nose. The heights 81 can also be determined. For this, step 207 of matching the 3D model of the pair of glasses with the pair of glasses 31 can be carried out. This allows to obtain precise contours of the pair of glasses 31 on image 2 and then measure the distances between the pupils and the edges of the pair of glasses 31. Subsequently, a virtual simulation of a wearing of the future pair of glasses will make it possible to transfer these values ​​to obtain the heights of the future pair of glasses by taking into account the differences between the heights of the two pairs of glasses (pair of glasses 31 worn and future pair of glasses). A future pair of glasses is understood to be a pair of glasses which will have been chosen by the user and for which we seek to determine the ocular measurements allowing the manufacture and the assembly of suitable corrective lenses. A wearing of a pair of glasses is understood to be the positioning of the pair of glasses on the face of the wearer.

[0107] A virtual pair corresponding to the future pair of glasses can be positioned on the face 4 thanks in particular to eyeglass distance values ​​and positioning values ​​on the nose learned by machine learning from a set of images of pairs of glasses worn corresponding to real situations, called ground truth, for which, in particular, the eyeglass distance values ​​and the positioning values ​​on the nose are known.

[0108] In the case of a glasses model constructed from frame marking, corrective parameters of the glasses model may have an influence on an interpretation of the frame marking. An interpretation of the frame marking means, for example, the way in which the frame marking is used when aligning the glasses with the glasses model constructed from the frame marking. These corrective parameters could be, for example, a curve of the pair of glasses, a lens-eye distance, or any other parameter having an influence on the measurement report. These corrective parameters can be estimated using machine learning.

Claims

Claims

1. A method (1) for determining eye measurement comprising the following steps: - select (100) an image (2) in a video stream, - identify (200) a known object (3) in the selected image, - detecting (300) a face (4) in the selected image (2), said face (4) comprising two sides, called left side and right side, each side comprising an eye (5) and an ear (6), - determining (400) an ocular measurement of the face using a reference length of the known object (3), measured in the selected image, characterized in that the image selection step comprises the following sub-steps, - detect (101) the face (4) in the video stream, - determining (106) for each side of the face (4) an ordinate of a first reference point (50) of the eye (5), called the eye ordinate, and an ordinate of a first reference point (60) of the ear (6), called the ear ordinate, - determine (107) for each side of the face (4) a length between a second reference point (51) of the eye (5) and a second reference point (61) of the ear (6), called eye-ear length, - selecting (110) the image (2) in the video stream so that for each side of the face (4), a difference between the eye ordinate and the ear ordinate is less than an ordinate threshold and so that a difference between the eye-ear length on one side of the face and the eye-ear length on the other side of the face is less than a length threshold.

2. Method according to claim 1, in which, for each side of the face (2), the first point (50) and the second reference point (51) of the eye are the outer corner of the corresponding eye (5) and the first point (60) and the second reference point (61) of the ear are the point of the corresponding ear (6) on which a branch will rest. of a pair of glasses.

3. A method according to any preceding claim, wherein the ordinate threshold and the length threshold are less than 2% of a face width.

4. Method according to one of the preceding claims, in which the step of selecting the image (2) comprises the following sub-step: - Displaying (108) the video stream with at least one visual indicator (7) each at least one visual indicator (7) being a function of at least one of the following: the right eye ordinate, the left eye ordinate, the right ear ordinate, the left ear ordinate, the right eye-ear length and the left eye-ear length.

5. Method according to one of the preceding claims, in which the sub-step of detecting (101) the face (2) in the video stream comprises the following sub-steps: - estimate (102) a 3D face model corresponding to the face, - matching (103) a position and orientation of the 3D face model with the face in the video stream.

6. Method according to claim 5, in which the sub-step of detecting (101) the face (2) in the video stream comprises the following sub-step: - refining (109) the estimation of the 3D face model from several refining images of the video stream, said refining images having different orientations of the 3D face model.

7. Method according to one of the preceding claims, in which the step of selecting (100) the image (2) comprises the following sub-steps: - determine (104) a face size, - determining (105) a center of the face, and in which the image (2) is selected in the video stream so that the size of the face is less than a size threshold and so that a distance between the center of the face and a center of the image is less than a centering threshold.

8. The method of claim 7, wherein the size threshold is less than 20% of an image size and the centering threshold is less than 2% of the face width.

9. Method according to any one of claims 7 to 8, in which the step of selecting (100) the image comprises the following sub-step: - Displaying (108) the video stream with at least one visual indicator (7) each at least one visual indicator (7) being a function of at least one of the following: face size, face center, right eye ordinate, left eye ordinate, right ear ordinate, left ear ordinate, right eye-ear length and left eye-ear length.

10. Method according to one of the preceding claims, in which the known object (3) is a pair of glasses (31) worn on the face by a user whose face is detected in the image and the video stream.

11. A method according to claim 10, wherein the pair of glasses (31) has been previously scanned.

12. A method according to claim 10, wherein a lens width and a bridge width of the pair of glasses (31) are previously known.

Images