What model to use for images with size 4k or above

[gluttonds-sys]

For images with a resolution of 4k or above, is it better to use 2xl or xl models for detection, or compress them to 1008 and enable multiple sizes? Is it appropriate to increase the resolution with xl and 2xl models

[isaacrob]

Resolution of input image is irrelevant, what matters is the size and subtlety of the objects. If your images always have one person in them, for example, running the nano at native resolution of 312 is probably sufficient regardless of how large the image is on disk :) generally I would recommend using larger sizes of the models instead of increasing resolution for smaller ones, and if that isn't sufficient to achieve target accuracy, then increase resolution of the largest ones. This is because the model architectures are discovered via NAS which considers resolution as part of the search space, so the larger models run at higher resolution than the smaller ones but also attempt to optimally scale up other parameters in the network. Just scaling resolution of the nano segmentation model, for example, misses out on the extra decoder layers and queries used by the larger sizes, while also running slower at higher resolutions due to the lesser number of windows.

[gluttonds-sys]

输入图像的分辨率无关紧要，重要的是物体的大小和细节。例如，如果你的图像中始终只有一个人，那么无论图像在磁盘上有多大，以 312 的原始分辨率运行 nano 模型可能就足够了 :) 通常，我建议使用更大尺寸的模型，而不是提高较小模型的分辨率；如果这样仍然无法达到目标精度，再提高最大尺寸模型的分辨率。这是因为模型架构是通过 NAS 算法发现的，而 NAS 算法会将分辨率视为搜索空间的一部分，因此较大的模型不仅比小的模型使用更高的分辨率，而且还会尝试优化网络中其他参数的缩放。例如，仅仅提高 nano 分割模型的分辨率，会忽略较大尺寸模型使用的额外解码器层和查询，同时由于窗口数量较少，在高分辨率下运行速度也会更慢。
如何加载和训练 xl 或 2xl 模型？

[isaacrob]

Requires the rfdetr_plus extension: pip install rfdetr[plus]. Note that those models require the PML license