足不出户 走遍全世界
A | 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。 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。 Monica Lewinsky, American activist and participant of a high-profile sex scandal in the 1990s involving the 42nd US President Bill Clinton, has launched a meme jab against another ex-POTUS, Donald Trump, following recent revelations about his alleged conduct during the US Capitol riots of 6 January 2021.,The revelations in question came in the form of a testimony delivered at the sixth public hearing of the January 6th Committee by Cassidy Hutchinson, former aide to Trump’s Chief of Staff Mark Meadows.,In her testimony, Hutchinson claimed that on 6 January 2021, Trump attempted to grab the steering wheel of the presidential vehicle he was traveling in that day, as he allegedly sought to join his supporters at the US Capitol while the head of his Secret Service detail, Bobby Engel, sought to return Trump back to the White House.,In response to this development, Lewinsky poked fun at the ex-POTUS by using a version of the so-called “distracted boyfriend” meme that contains reference to the infamous tape where Trump can be head bragging that, “when you are a star”, you can supposedly “do anything” to women, including grabbing them “by the p*ssy”.,(i think i did this correctly) pic.twitter.com/ZzTSlTNhMm— Monica Lewinsky (she/her) (@MonicaLewinsky) June 28, 2022,Quite a few social media users appeared rather amused by Monica’s stunt, praising her over it, while others used their opportunity to joke about Trump.,“If his hands were bigger, he might have succeeded in making it to the Capitol,” one netizen suggested.,“This may be the single greatest takedown in the history of the internet,” another tweeted.,Donald Trump himself has already denied Hutchins’ allegations, and Engel and the agent who drove the presidential SUV on that day are also reportedly willing to testify that before the US House Select Committee that Trump did not “lunge at the steering wheel”.。 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。论文称，部分输出细节甚至超过训练时 Blender 渲染结果，可保留猫胡须和单根发丝边缘。 性能方面，论文给出两组处理时间数据：小模型处理 81 帧视频约需 6 秒；大模型参数量为 140 亿，处理同样长度视频约需 10 秒。
B | 参考。
Current article:http://www.liawucunsainantaonao.sbs/news/20260826_7622717.xls
Published on:04:09:08
推荐阅读
