ECCV 2026 Paper Introduces MegaKPT Dataset with 1.3M Instances and DINOv3-based GKDT Model
NielsRogge · x · 2026-08-05
This research, slated for ECCV 2026, introduces MegaKPT, a large-scale unified keypoint dataset comprising over 1.3 million diverse object instances from 29 existing datasets, featuring high-quality unified annotations and keypoint text descriptions.
Leveraging this dataset, the authors trained GKDT (General Keypoint Detection Transformer), a DINOv3-based model that supports visual prompts, text prompts, or both. It demonstrates strong performance and generality across broad categories, achieving over 90% PCK@0 in most categories.
More from Research
- Why Robotics Is Hard: The 'Fall Damage' of Sim2Real — generativist · 2026-08-05
- Research: Using LLMs to Automate Theory Building in Cognitive Science — JessicaHullman · 2026-08-05
- 'The Hundred-Page Language Models Book': Build LLMs from Scratch with PyTorch — burkov · 2026-08-05
- Perfect Data QA is the Key Unlock for Recursive Self-Improvement — simonguozirui · 2026-08-05
- AI Writes Its First Complete Genome, Treating DNA as a Natural Language — latkins · 2026-08-05
- Why Self-Improving Autoresearch Loops Fail: Paper Reveals Sharp Drop in LLM Judge Accuracy — dair_ai · 2026-08-05