Grigorios-Aris Cheimariotis

3D reconstruction for large scale indoor and outdoor environments constituted a challenging problem in inverse graphics and computer vision research community. The scalability issue that concerns only the geographic dimensions of the reconstructed scene in terms of scale is a inevitable bottleneck. This could be confronted by state-of-the-art methods which a comparative study should be reveal new insights of the digital environment representation. In this paper, a comparative study is facilitated including few literature methods with accordance by common evaluation metrics and computational complexity. A novel loss function referring to the scene details is also discussed by accessing the density variable of the NeRF multilayer perceptron output as alternative research direction. To this end, it presents efficiency advantange in terms of robustness to noise, scale and sentivity to distortions combined with sufficiency on accessed GP-GPUs memory and footprint in several datasets while it was validated and compared with state-of-the-art Inf-NeRF, Grid-NeRF and Nerfacto techniques.

This work proposes HAL-NeRF v2, a localization pipeline that couples direct pose regression with Monte Carlo–based refinement on neural scene representations. Building upon the original HAL-NeRF framework, the proposed system leverages Gaussian Splatting for fast, high-detail synthetic view generation during pose regressor training, and NeRFs for efficient pose refinement. The refinement stage is redesigned with Cauchy loss, systematic resampling, and maximum-likelihood estimation,instead of the HAL-NeRF steps, to improve convergence stability in scenes with transient elements and/or incomplete mapping. In addition, HAL-NeRF v2 evaluates multiple rendered views in parallel during refinement, allowing the particle filter to rapidly disambiguate candidate viewpoints. The experimental results in the Cambridge Landmarks dataset demonstrate an accuracy equivalent to that of HAL-NeRF v1 (0.09 m / 0.61° median translational/rotational error) while achieving a 25× reduction in run time.

Intravascular Optical Coherence Tomography (IVOCT) images provide important insight into every aspect of atherosclerosis. Specifically, the extent of plaque and its type, which are indicative of the patient’s condition, are better assessed by OCT images in comparison to other in vivo modalities. A large amount of imaging data per patient require automatic methods for rapid results. An effective step towards automatic plaque detection and plaque characterization is axial lines (A-lines) based classification into normal and various plaque types. In this work, a novel automatic method for A-line classification is proposed. The method employed convolutional neural networks (CNNs) for classification in its core and comprised the following pre-processing steps: arterial wall segmentation and an OCT-specific (depth-resolved) transformation and a post-processing step based on the majority of classifications. The important step was the OCT-specific transformation, which was based on the estimation of the attenuation coefficient in every pixel of the OCT image. The dataset used for training and testing consisted of 183 images from 33 patients. In these images, four different plaque types were delineated. The method was evaluated by cross-validation. The mean values of accuracy, sensitivity and specificity were 74.73%, 87.78%, and 61.45%, respectively, when classifying into plaque and normal A-lines. When plaque A-lines were classified into fibrolipidic and fibrocalcific, the overall accuracy was 83.47% for A-lines of OCT-specific transformed images and 74.94% for A-lines of original images. This large improvement in accuracy indicates the advantage of using attenuation coefficients when characterizing plaque types. The proposed automatic deep-learning pipeline constitutes a positive contribution to the accurate classification of A-lines in intravascular OCT images.