A mobile manipulator consists of a mobile platform and a robot arm, which can per-form multiple tasks over a wide workspace. However, due to the navigation perfor-mance and localization errors of a mobile platform, pose errors occur even when the target pose required for the target task is pre-taught. To address this issue, a camera is attached to the end-effector of the robot to recognize the pose required for the task through marker detection from the image. At this point, there is a kinematic error be-tween the end-effector frame and the attached camera frame, so the calibration be-tween the two frames is needed to reduce the error. However, due to the characteristics of the image-based method, even if the calibration is performed, there is still a pose error in the process of calibration and marker detection due to camera distortion. To address these issues, this study produces a deep neural network that compensates for errors in the estimated marker pose and errors in the coordinate transformation matrix between the end-effector frame and the camera frame, and then proposes a deep learn-ing-based pose estimation method that minimizes three-dimensional pose errors through this neural network. The proposed method reduced the pose error by 62.4% compared to the previous method and showed 100% success in 50 pick-and-place tasks in which the target object is placed to the target jig with a tolerance of 1mm.