Dense 3D reconstruction from RGB images typically assumes fixed camera positions, even for mobile devices. However, this assumption doesn’t apply when poses are dynamic (e.g., updated through bundle adjustment and loop closure). While this has been addressed for RGB-D, the RGB-only setting is largely unexplored. We propose to address this gap.

Researchers from MIT have developed a new method called CONSENSUS GAME to improve language model (LM) decoding processes. It combines generative and discriminative approaches to extract the best estimate of truth from contradicting signals. The game-theoretic method, known as EQUILIBRIUM-RANKING, outperformed existing techniques in question-answering benchmarks. This research demonstrates how game theory can enhance coherence…

Two friends, Salvatore Aiello and Monica Powers, met at an online event and created an AI tool called DimeADozen. They spent $185 to make it and sold it for $150,000. Even after selling it, they continue to help the tool grow. Lesson learned: small ideas can turn into something big!

NYC Mayor Eric Adams is using AI-generated deepfake technology to make automated robocalls to his city’s residents. The AI creates audio of Adams speaking in various languages, allowing him to reach a wider audience. While practical, this raises ethical concerns as the recipients are not informed that the voice they hear is AI-generated. Some criticize…

IBM Research’s lab in California has developed the NorthPole chip, which could accelerate the realization of AI on the edge. The chip, with 256 cores and its own memory, eliminates the need for off-chip memory, making it faster and more energy-efficient. IBM’s initial results indicate that NorthPole outperforms other chips in terms of power usage,…

Researchers have developed OmniControl, a diffusion-based human generation model that incorporates spatial control signals over any joint at any given time. This model addresses the limitations of previous techniques in integrating variable spatial control signals, allowing for more realistic and precise human motion. By combining spatial and realism guidance, OmniControl achieves better motion realism and…

Researchers from the Australian National University, the University of Oxford, and the Beijing Academy of AI have developed an AI system called “3D-GPT” that can generate 3D environments based on text prompts. The system breaks down complex tasks into segments handled by specialized AI agents, resulting in streamlined 3D asset creation. Although not yet photorealistic,…

Researchers are finding that ChatGPT, OpenAI’s advanced language model, can provide useful feedback as an alternative to human reviewers in the peer review process. In a study, over 50% of ChatGPT’s comments on Nature papers and over 77% on ICLR papers aligned with human reviewers’ points. However, concerns about bias and accuracy persist, and it…

Retrieval Augmented Generation (RAG) has revolutionized open-domain question answering by using a retrieval module to find relevant context passages and a generative module to provide answers. However, vector search, one of the critical components, has limitations in capturing nuanced reasoning, handling complex questions, and modeling diverse relationships. Knowledge graph prompting, which encodes various connections into…

Summary: The French association Data for Good released a white paper examining the environmental impact of language models. ChatGPT’s monthly usage emits 10,000 tons of CO2, equivalent to 0.1% of the yearly carbon footprint of individuals in France/UK. If ChatGPT+ with GPT-4 is used, the carbon footprint could increase by 10 to 100 times, contributing…