EVENT DOSSIER
The XDG model fine-tunes Depth Anything 3 using a lightweight LoRA adapter, reusing the camera token as a paired classification token and utilizing its cross-view geometric reasoning capabilities for visual disambiguation. This approach avoids the additional computational overhead of heavy decoders. Experiments show that XDG remains competitive with state-of-the-art methods in paired and reconstruction benchmarks, while providing more than three times faster inference speeds; in a single LaMAR scenario involving thousands of images, it can save over 10 hours of time spent on visual disambiguation. The related code is now open-source.
arXiv:2608.29733v2 XDG 提出高效视觉去重模型，推理速度提升 3 倍
XDG 模型通过轻量级 LoRA 适配器微调 Depth Anything 3，将相机令牌重用作成对分类令牌，实现了视觉去歧义的高效处理。该方法利用 3D 基础模型固有的跨视图几何推理能力，直接适配骨干网络表示，避免了重型解码器的额外计算开销。实验表明，XDG 在成对和重建基准测试中与最先进方法保持竞争力，同时提供超过 3 倍的推理速度提升；在包含数千张图像的单个 LaMAR 场景中，可节省超过 10 小时的视觉去歧义处理时间。代码已开源。