Accurate attitude estimation is essential for stable guidance and control during rocket recovery, yet it remains challenging because the target undergoes rapid pose changes, occupies only a limited image area over a large observation corridor, and often exhibits weak texture and approximate axial symmetry. To address these issues, this paper proposes a large-field binocular-vision-based attitude determination method for rocket recovery. First, a distortion-aware stereo calibration strategy based on stitched control points is developed to enable reliable geometric modeling over a large measurement field with a portable calibration target. Second, a robust contour extraction pipeline is constructed by combining bilateral filtering, gradient enhancement, and multi-threshold Canny fusion. Third, the rocket central axis is reconstructed by fitting 2D midlines in rectified stereo images and intersecting their corresponding back-projection planes, which improves stability over point-wise triangulation. Finally, pitch and yaw are derived from the recovered 3D axis direction, while roll is estimated by phase correlation on the polar-unwrapped base image under a temporal continuity constraint. Experiments on a 1:20 cylindrical scale model show RMS reprojection errors of 0.056 px and 0.066 px for the left and right cameras, respectively, and a 3D checkpoint RMSE of 33.42 mm. On a 100-frame sequence, the proposed method achieves RMSEs of 1.58°, 1.54° and 1.41° for roll, pitch, and yaw, respectively, outperforming ORB+PnP, SGBM, and Chamfer-based baselines. The results demonstrate that the proposed method provides an accurate and practical optical solution for external attitude measurement in rocket-recovery scenarios.