The Latent Space LIVE mini-conference at NeurIPS 2024 showcased the latest breakthroughs in computer vision. Speakers from Roboflow and Moondream discussed significant trends, such as the transition from per-image to video processing seen in models like Sora and SAM2. They also noted how DETR models are now outpacing YOLO in real-time object detection. Additionally, they addressed the challenges faced by large language models (LLMs) in capturing intricate visual details, referencing important research like MMVP, Florence 2, PolyGemma 2, and AIM-V2. The session included a lively discussion on the shortcomings of current vision-language models (VLMs) in complex visual tasks and explored how synthetic data and chain-of-thought prompting could enhance their effectiveness.
