Posture estimation is a significant area of research in the realm of computer science. The objective is to precisely determine human posture information by conducting computer analysis on photos or videos of the human body. Posture estimation has made considerable advancements and advances because to the increasing popularity of deep learning technologies. The paper presents a robust and accurate technique, based on deep learning, for estimating the 3D pose of fish without the need for markers. Initially, through the analysis of labels in 2D posture estimation and the assessment of the effectiveness of various neural network variations, an ideal 2D pose estimation model is achieved. The 3D posture estimation utilizes the Bundle Adjustment optimization approach and the precise key point triangulation methodology for camera calibration. The 3D pose estimate is achieved by a multi-dimensional optimization method, both at the 2D and 3D levels. The simultaneous presentation of 2D and 3D orientation from various viewpoints demonstrates that this approach may effectively accomplish the estimation of fish’s 3D orientation. Furthermore, we also compute several kinematic properties of fish, which serves as a basis for future improvements in pose estimation accuracy.