Visual Model

On September 20th, Goldman Sachs announced that TrafficVLM (Note: Traffic Visual Language Model) would help enable users to achieve global traffic control and enhance driving experience. It is described that in modern transport environments, drivers often face the challenge of information-blind areas: when they travel through complex traffic, they can only see the currents in front of them, but they cannot predict which lanes 100 metres away are about to be blocked; and when they travel on the freeway, it is difficult to anticipate the “grows of ghosts” triggered by a slight brake in the forward. These limitations of a local perspective make it difficult for drivers to make optimal decisions. So, Tr..

DeepSeek's official public website released a blog post yesterday (December 13), announcing the open source DeepSeek-VL2 model, which has achieved very advantageous results in various evaluation indexes, and officially said that its visual model has officially entered the era of Mixture of Experts (MoE). Citing the official press release, 1AI attached the highlights of DeepSeek-VL2 as follows: Data: double the quality of training data compared to the first generation of DeepSeek-VL, and introduction of terrain understanding, visual localization, visual storytelling...

Researchers from Apple and the Swiss Federal Institute of Technology in Lausanne (EPFL) have jointly developed a single any-to-any modality model that can be trained on dozens of highly diverse modalities and co-trained on large-scale multimodal datasets and text corpora. The model, named 4M-21, is trained on 21 different modalities and completes at least 3 times more tasks than existing models without losing performance. The study used the 4M pre-training scheme to expand the scale of the model and dataset, increase the type and number of modalities involved in training the model, and train on multiple datasets...

According to Meta's official press release, the company has developed a visual model called "SceneScript", which claims to be able to use a programmable language to quickly "build" scenes, infer room geometry in real time, and convert related data into architectural approximations. Image source Meta's official press release Meta claims that the relevant method can efficiently and lightly build indoor 3D models, claiming that "only a few KB of memory are needed to generate clear and complete geometric shapes", and the relevant shape data is "interpretable", and users can easily read...