Xinyi Ying

SIATD中有标注的1-50序列作为测试集，剩下的250段为训练集。测试集还使用了空天杯v1和Anti-UAV的数据。具体内容详见论文。 论文中对应部分在IV. EXPERIMENTS；A. Experiment Settings；1) Datasets “In this paper, we employ the 1st − 50th sequences with target annotations of SAITD as the test datasets and the remaining 300 sequences...

MoCoPnet做的是图像超分辨，不是小目标检测，所以训练不需要标注。 3090训练100K用时约24h，具体时长看工作站性能

可能是硬件导致的数据读写比较慢？可以试试使用固态硬盘存储数据。

Win10 is not friendly for compile, we recommand ubuntu for D3D compile. If you have to compile D3D for win10, we have to prepare Visual Studio for compiling files based...

You have to compile D3D before using it. See "Build" in readme.md

We scale up the kernel sizes to 5x5x5, 7x7x7, and can successfully run this code. Details are as follows: 

Try to set the groups to 1.

The current version of D3D can work under CUDA11.7 on Ubuntu. You have to compile D3D first before training and test. Please refer to [README.md](https://github.com/XinyiYing/D3Dnet/blob/master/README.md) for more details.

You can choose any dimension you want to deform by setting parameter dimension. See ```code/dcn/test.py``` for more details. Example codes of deform among three dimension: dcn = DeformConvPack_d(inC, outC, kernel_size=[kT,...

We recommend users to use ubuntu for compiling. Many users who use win for compiling always report a variety of environmental bugs, including the ninja error.