Method and apparatus with pose estimation

A processor-implemented method, comprising:

• obtaining, using a depth network, a respective depth image for each of a plurality of successive input images;
• obtaining, using a pose network, respective image pose transformation matrices, and respective inverse image pose transformation matrices, between images, of the successive input images, at adjacent time points;
• calculating a composite loss value based on respective comparisons for each time point of a corresponding input image and a composite image generated using a corresponding inverse image pose transformation matrix applied to a successive image of the corresponding image, and respective comparisons for each time point of the successive image and a composite image generated using a corresponding image pose transformation matrix applied to the corresponding image;
• calculating a pose loss value based on a result of comparing image position information, obtained from a result of an accumulating of the image pose transformation matrices, and sensor position information obtained from a sensor;
• and
• updating training of the depth network and the pose network based on a total loss value calculated from the composite loss value and the pose loss value.

An apparatus, the apparatus comprising:

• a position sensor;
• and
• one or more processors configured to:
• obtain, using a depth network, a respective depth image for each of a plurality of successive input images;
• obtain, using a pose network, respective image pose transformation matrices between images, of the successive input images, at adjacent time points;
• obtain, based on initial pose information and the respective image pose transformation matrices, image pose information for each of the adjacent times;
• estimate final pose information dependent on the obtained image pose information for each of the adjacent times;
• accumulate the image pose transformation matrices;
• and
• calculate a pose loss value based on a result of comparing image position information, obtained from a result of the accumulating, and sensor position information obtained from the position sensor.

A vehicle, the vehicle comprising:

• a position sensor;
• and
• one or more processors configured to:
• with respect to each of a plurality of position samplings by the position sensor, repeat:
• an obtaining, using a pose network, of respective image pose transformation matrices, and respective inverse image pose transformation matrices, between images of a plurality of input images;
• an obtaining of image pose information dependent on the respective image pose transformation matrices;
• a calculating of a composite loss value dependent on the respective image pose transformation matrices, the respective inverse image pose transformation matrices, the plurality of input images, and respective depth images for the plurality of input images;
• an obtaining of image position information dependent on a result of an accumulating of the respective image pose transformation matrices;
• and
• a calculation of a pose loss value dependent on the obtained image position information and sensor position information from the position sensor;
• and
• update training of the pose network dependent on a result of one of the repeated calculations of the composite loss value and the pose loss value for a corresponding one of the plurality of position samplings.