The following are the most intriguing scientific publications from the past year. It is a selected list of the most current achievements in artificial intelligence and data science, chronologically grouped with a link to a more detailed article and source code (if applicable).
ShaRF
Consider how great it would be to snap a photograph of an object and have it rendered in 3D for use in a film or video game you’re developing or in a 3D scenario for an illustration.
The researchers demonstrate a technique for predicting neural scene representations of items from a single image. The method’s central idea is to estimate a geometric scaffold for the item and use it as a guide for reconstructing the object’s underlying radiance field. Their formulation translates a latent code to a voxelized shape and then renders it an image, with the object’s appearance controlled by a second latent code. They optimize the latent codes and the networks during inference to fit a test image of a new item. Their model tune a single image due to the explicit disentanglement of shape and appearance. They can then generate unique geometrically consistent views and adequately depict the input object.
Additionally, their technique generalizes to images not part of the training domain (more realistic renderings and even real photographs). Finally, the inferred geometric scaffold precisely represents the object’s three-dimensional shape. The researchers demonstrate the success of their approach in various studies using both synthetic and real photos.
Thinking fast and slow in AI
This article presents a study strategy for advancing artificial intelligence inspired by cognitive theories of human decision making. By gaining insight into the underlying causes of various human capabilities that are still lacking in AI (for example, adaptability, generalizability, common sense, and causal reasoning), we may integrate these causal components into an AI system and get analogous capabilities. Furthermore, the researchers hope that the high-level description of their vision will spur the AI research community to define, experiment with, and evaluate novel methodologies, frameworks, and evaluation metrics to understand both human and machine intelligence better.
Automatic detection and quantification of floating marine macro-litter in aerial images
The dangers posed by human floating marine macro-litter (FMML) to marine species and ecosystems, in general, are well recognized. Worldwide, dedicated monitoring programs and mitigating measures are in place to address this issue, with the support of new technology and the automation of analytical processes rising.
The current study created methods for recognizing and measuring FMML in aerial pictures and a web-based application that enables users to remember FMML in sea surface photographs. The suggested algorithm for a deep learning technique uses convolutional neural networks (CNNs) that are capable of learning from unstructured or unlabelled material. The CNN-based deep learning model trains using 3723 aerial photos (50 per cent of which contained FMML and 50 per cent of which did not) captured by drones and planes over the waters of the NW Mediterranean Sea. Image classification (using all photos for training and testing the model) and cross-validation (using 90% of images for training and 10% for testing) had accuracy values of 0.85 and 0.81, respectively. After that, the researchers used the Shiny package in R to create a user-friendly application for identifying and quantifying FMML in aerial pictures. The application of this and comparable algorithms significantly reduces the time required for FMML detection and quantification, assisting in monitoring and assessing this environmental danger. However, automatic FMML monitoring in the open sea remains a technological barrier, and additional study is necessary to increase the accuracy of existing algorithms.
Source: indiaai.gov.in
