Despite the elegance of the mathematics, these epipolar techniques do not address some pertinent practical issues in the Structure from Motion problem. In particular, they are reliable for perfect features and images with wide baselines but are sensitive to noise [55]. The formalism focuses on linear reformulations and only considers 2D measurement errors as an after thought. Thus it can exhibit numerical instabilities. These are especially evident when the baseline (i.e. relative camera translation between frames) is small. The case of no translation and pure rotation in the motions are actually degenerate cases for epipolar geometry. As the camera configurations approach degeneracies, the epipolar results vary wildly and noise causes numerical ill-conditioning. In addition, in the trifocal tensor case, the intermediate computation of higher order polynomials could also be prone to noise sensitivity.
Essentially, these techniques focus on and perform best in 2-frame and 3-frame structure from motion with large baseline and small 2D measurement error. Their application to image sequences involves further processing and often requires some manual supervision. For instance, Faugeras discusses the special treatment that the trifocal technique requires on image sequences in post-production type applications [17]. Herein, a human user must preselect the triples of appropriate frames in a sequence to guarantee a wide range of camera motion (i.e. wide baseline) in the trifocal tensor calculations. In addition, the technique does not use all the images in a sequence, only appropriate triples. Therefore, the estimates of structure and motion are combined together using a somewhat unprincipled interpolation calculation. Since not all image frames are used and interpolation is relied on (be it linear or higher order), this technique discards data and hence compromises some accuracy.