Optimal detection methods for small targets rely on whitened matched filters, which convolve the measured data with the signal model, and whiten the result with the noise covariance. In real-world implementations of such filters, the noise covariance must be estimated from the data, and the resulting covariance estimate may be corrupted by presence of the target. The resulting loss in SNR is called 'target capture'. Target capture is often thought to be a problem only for bright targets. This presentation shows that target capture also arises for dim targets, leading to an SNR loss which is independent of target strength and depends on the averaging method used to estimate the noise covariance. This loss is due to a 'coherent beat' between the true noise and that portion of the estimated noise covariance due to the target. This beat leads to 'ghost targets', which diminish the target SNR by producing a negative target ghost at the target's position. A quantitative estimate of this effect will be given, and shown to agree with numerical results. The effect of averaging on SNR is also discussed for data scenes with synthetic injected targets, in cases where the noise covariance is estimated using 'no target' data. For these cases, it is shown that the so-called 'optimal' filter, which uses the true noise covariance, is actually worse than a 'sub-optimal' filter which estimates the noise from scene. This apparent contradiction is resolved by showing that the optimal filter is best if the same filter is used for many scenes, but is outperformed by a filter adapted to a specific scene.