河神

谷歌 DeepMind 发布 GenCeption：打破计算机视觉桎梏_我的网站

不要恋爱要结婚

A | London, Oct 19 (UNI) The resignation of Liz Truss from the post of UK prime minister will not solve the problems in the country, Foreign Secretary James Cleverly told Sky News on Wednesday.

Cleverly also noted that he can understand people who are dissatisfied with the government's policies, but removing Truss from her post is not an option. In his opinion, people who demand the resignation of the current prime minister do not have a plan for further action in the same way as was the case with the resignation of her predecessor, Boris Johnson.

On Tuesday, a YouGov poll showed that only 10% of surveyed Britons view Truss favorably after her ineffective economic plan, while another survey by the same pollster revealed 55% of the Conservative party members want Truss to resign.

The Truss administration has been harshly criticized for its economic policy that proved unable to curb inflation and price hikes, as well as sparked concerns about an increase in public debt. The so-called mini-budget was presented on September 23 by then Chancellor Kwasi Kwarteng, who resigned last Friday.

After the plan's announcement, the yield on five-year UK government bonds rose to its highest level since 2008, at 4.6%, which meant a decrease in demand for debt securities. Consequently, sterling fell to an all-time low of $1.054 per pound.

UNI/SPUTNIK GNK。 IT之家 7 月 21 日消息，科技媒体 The Decoder 昨日（7 月 20 日）发布博文，报道称谷歌 DeepMind 发布 GenCeption 模型，将预训练的视频生成器重新用于深度估计和分割等经典计算机视觉任务。 IT之家援引博文介绍，大语言模型在学习预测下一个 Token 的时候，在训练过程中往往需要吸收语法、世界知识和上下文关系等内容。

B | 但是在计算机视觉领域，视觉模型缺少等效的训练方法，主要由专业模型主导，包括用于分割的“Segment Anything”和用于深度估计的“Depth Anything”，每个模型都使用其特定的架构。谷歌 DeepMind 团队为此提出 GenCeption 模型方案，尝试将一个“生成视频”的 AI 模型逆向改造成一个能“理解世界”的视觉分析引擎。 GenCeption 打破了传统计算机视觉“一个任务一个专用模型”的格局，仅凭单一模型就能同时做好深度估计、图像分割、3D 姿态估计、表面法线预测和相机姿态估计等核心视觉任务。 GenCeption 基于阿里巴巴开源视频模型通义万相 Wan2.1 系列训练，与传统扩散模型需多步去噪不同，GenCeption 在一次前向传播中完成预测，从而提升视觉任务处理速度。模型通过文本提示指定任务，可输出深度图、表面法线图、分割掩码，并可处理相机运动表示。训练数据以合成为主。论文称，数据集仅包含 7500 段视频，由 800 个数字人体模型与 200 段动作捕捉序列组合生成，再通过 Blender 在不同背景和镜头角度下渲染。泛化方面，GenCeption 几乎只在单人合成视频上训练，但可处理真实多人视频，也可迁移到动物和类人机器人类别。