Posts by Larsen D'hiet

Trellis 2: Scaling 3D Generation with Improved Efficiency and Control

In a year marked by rapid advances in 3D generative modeling, Trellis 2 makes for one of the most exciting architectural updates this year. It introduces Omni-Voxels, a native 3D representation that encodes geometry and PBR materials directly in aligned 3D space. Combined with the new Sparse Compression VAE, this enables more efficient compression of very high-resolution assets at improved inference speeds.

Read Article

Why DETRs are replacing YOLOs for real-time object detection

Detection Transformers (DETRs) have matured into real-time capable object detectors, rivaling YOLOs in both speed and accuracy. Despite early challenges, advancements like deformable attention, denoising training, and top-k query selection paved the way for the first real-time Detection Transformer RT-DETR, introduced by a team of Baidu researchers in 2024. Recent innovations like D-Fine’s fine-grained localization and DEIMv2’s foundation-model backbones push accuracy even further. Additionally, all DETR models and weights are released under the permissive Apache 2.0 License, enabling free use and commercial adaptation. At Datameister, we integrate these cutting-edge models into our vision library for high-performance, adaptable, and production-ready detection systems for complex, specific problems.

Read Article

Datameister @ECCV 2024: Building a foundation

Datameister's presence at ECCV 2024 in Milan was a deep dive into the latest advancements in Computer Vision. This blogpost explores groundbreaking techniques like 3D Gaussian Splatting and its enhancements, such as WildGaussians and Gaussian Frosting, showcasing their potential to revolutionize real-time rendering and scene representation.

Read Article