Terry Zhang
Terry Zhang
Thanks for your question. It is randomly selected in most of our experiments. For JVTC, a fixed txt file is needed. Therefore, we also provide an example txt file containing...
Hi, you can download the pretrained model on BaiduNetdisk or beihang netdisk.
Sorry, I think this plugin is no longer needed to run the program.
是的，没错
Hi. Sorry, we have only evaluated our synthesized data with basic loss functions used in CBN. The triplet loss or circle loss you mentioned may be effective for the unreal...
> @FlyHighest 同学你好，感谢你公开的数据集。请问一下你给的list_train_label.txt是处理后的label吗？我只需要load你txt里面的label就行了是吧？ @michuanhaohao 是的，可以读出来直接用，只是随机sample了部分数据写成了这个txt文件，您也可以根据需要的数据量自己生成。 ``` with open(txt_path) as f: line = f.readlines() self.img_list = [os.path.join(dataset_dir, i.split()[0]) for i in line] self.label_list = [int(i.split()[1]) for i in line] self.cam_list =...
@michuanhaohao 是一致的，从头训练unreal我们也没做过太多实验，效果不如imagenet->unreal->other datasets好
> 在真实数据上测试模型，且都在imagenet上进行模型权重初始化，真实数据和虚拟数据混在一起训练和拿虚拟数据pretrain然后在真实数据上finetune相比，哪种方式更好？作者有试过类似的尝试吗？ @AlphaPlusTT 谢谢您的提问。训练方式1：imagenet->unreal->target ; 训练方式2：imagenet-> unreal+target. 这两种训练方式，在target上测试的结果，方式1更优。但是训练出来的模型在其他非target数据集测试的话，方式2更优。
Hi, I don't have the accurate file list. How did you randomly take images? Please try to follow the experiment instructions about CBN and use all sub datasets.
unrealv1.2 has only one scene and one person type. It's not a surprise that only using one subset doesn't show the advantages of unreal synthesis.