For 3-D reconstruction from uncalibrated sequences, projective reconstruction is needed as a preceding step, which is classified into a merging method and factorization. Factorization-based methods generally suffer less from drift and error accumulation than sequential merging. However, they assume that most corresponding points must remain in all frames. To overcome this limitation, we present a new frame grouping method for factorization-based projective reconstruction over long image sequences. The proposed method breaks a full sequence into subsequences by considering a survival rate of the corresponding points over a sequence, the reprojection error, and the distribution of matching points on the image. All projective reconstructions in subsequences are registered into the same coordinate frame for a complete description of the scene. Experimental results demonstrate that our algorithm can recover more precise 3-D structure than the sequential merging method.