Videos of sports injuries can potentially provide valuable information on non-contact ACL injuries. However, only the un-validated simple visual inspection approach has, so far, been used. Therefore, the purpose of this study was to test the accuracy and precision of researchers in estimating kinematics from video sequences of situations that typically lead to ACL injuries. We also tested if accuracy and precision could be improved through a training program. Using a routine surface marker based infrared, 240 Hz, 3D motion analysis system, we recorded running and cutting trials from three test subjects. Six observers were asked to provide estimates of kinematic variables from 27 video composites from one, two or three ordinary cameras, systematically varying viewing angles and time point of analysis. The observers thereafter went through a training program where 35 similar composites were analyzed, and feedback on the kinematics, as measured by the 3D motion analysis system, was provided on a group basis. Finally, the test was repeated to assess accuracy and precision. The mean error for knee flexion was -19 degrees, indicating a consistent underestimation. Hip flexion was underestimated by 7 degrees, but the standard deviation between the observers was 18 degrees on average, indicating poor consistency. Substantial errors were also found in the accuracy and precision of the other estimates. Only small group effects were seen from our training program. Based on these findings, results from studies using a simple visual inspection approach to describe joint motion must be interpreted with caution.