<p>Baseball is one of the most popular sports worldwide, in which pitching velocity and control, collectively referred to as pitch quality, are key factors influencing the outcome of a game. With the continuous collection of high-quality video data from baseball games, it has become increasingly feasible to reconstruct 3D pitching motions using pose estimation techniques. However, accurate motion reconstruction is hindered by challenges such as motion blur, particularly when the ball is thrown at high speed toward the strike zone. To address this issue, we propose a two-stage learning approach, consisting of motion imitation and refinement stages, to sequentially reconstruct baseball pitching motions from imperfect pose estimates extracted from video. Our framework combines physics-based simulations and deep reinforcement learning (DRL), along with domain-specific rewards, to guide the simulated character in throwing the ball to the target location at the desired speed while imitating the estimated pose and preserving the pitching style. We demonstrate the effectiveness of our framework through various experiments. Our method successfully reconstructs plausible, physically consistent pitching motions that closely resemble the original video, even when pose estimates are imperfect.</p>