We propose SelfSplat, a novel 3D Gaussian Splatting model designed to perform pose-free and 3D prior-free generalizable 3D reconstruction from unposed multi-view images. These settings are inherently ill-posed due to the lack of ground-truth data, learned geometric information, and the need to achieve accurate 3D reconstruction without finetuning, making it difficult for conventional methods to achieve high-quality results. Our model addresses these challenges by effectively integrating explicit 3D representations with self-supervised depth and pose estimation techniques, resulting in reciprocal improvements in both pose accuracy and 3D reconstruction quality. Furthermore, we incorporate a matching-aware pose estimation network and a depth refinement module to enhance geometry consistency across views, ensuring more accurate and stable 3D reconstructions. To present the performance of our method, we evaluated it on large-scale real-world datasets, including RealEstate10K, ACID, and DL3DV. SelfSplat achieves superior results over previous state-of-the-art methods in both appearance and geometry quality, also demonstrates strong cross-dataset generalization capabilities. Extensive ablation studies and analysis also validate the effectiveness of our proposed methods. Code and pretrained models are available at https://gynjn.github.io/selfsplat/.

Stack filters are a class of nonlinear filters with excellent properties for signal restoration. Unfortunately, present algorithms for designing stack filters can only be used for small window sizes because of either their computational overhead or their serial nature. This paper presents a new adaptive algorithm for determining a stack filter that minimizes the mean absolute error criterion. The new algorithm retains the iterative nature of many current adaptive stack filtering algorithms, but significantly reduces the number of iterations required to converge to an optimal filter. This algorithm is faster than all currently available stack filter design algorithms, is simple to implement, and is shown in this paper to always converge to an optimal stack filter. Extensive comparisons between this new algorithm and all existing algorithms are provided. The comparisons are based both on the performance of the resulting filters and upon the time and space complexity of the algorithms. They demonstrate that the new algorithm has three advantages: it is faster than all other available algorithms; it can be used on standard workstations (SPARC 5 with 48 MB) to design filters with windows containing 20 or more points; and, its highly parallel structure allows very fast implementations on parallel machines. This new algorithm allows cascades of stack filters to be designed; stack filters with windows containing 72 points have been designed in a matter of minutes under this new approach.

The theory of optimal stack filtering has been used in the difference of estimates (DoE) approach to the detection of intensity edges in noisy images. The DoE approach is modified by imposing a symmetry condition on the data used to train the two stack filters. Under this condition, the stack filters obtained are duals of each other. Only one filter must therefore be trained; the other is simply its dual. This new technique is called the symmetric difference of estimates (SDoE) approach. The dual stack filters obtained under the SDoE approach are shown to be comparable. This allows the difference of these two filters to be represented by a single equivalent edge operator. This latter result suggests that an edge operator can be found by directly training a (possibly nonpositive) Boolean function to be used on each level of the threshold decomposition architecture. This approach, which is called the threshold Boolean filter (TBF) approach, requires less training time but produces operators that are less robust than those produced by the SDoE approach. This is demonstrated and interpreted via comparisons of results for natural images.

This paper considers speech enhancement where the speech signal is modeled with hidden filter models (HFMs), when only noisy speech signal are available. The HFM is a parametric approach for representing the speech waveform in the time domain. We apply the nested EM algorithm for jointly estimating the clean signal and the parameters of HFM and noise model. A computationally efficient implementation of the algorithm is developed by the log-likelihood gradient based on the Kalman filter output in the estimation process. The resulting algorithm does not need framing and may be viewed in the time domain context.