WHHHHY
commented
2 months ago 我正在尝试仅在搅拌机生成的合成数据（渲染数据）上实现 PVN3D 代码。我只提供了合成图像，而不是真实图像（从渲染文件夹中生成的 pkl 文件中提取的参数和属性）。令我惊讶的是，我在训练中获得了 0 的改进，即在检查点文件夹中创建单个初始 .pth 文件后损失值停止。谁能告诉我为什么如果我只使用合成数据或者代码本身只有在获得一些真实图像时才能改进，训练不会得到改善？ 欢迎对此问题提出任何评论
Did you solved this problem?And how to do?
I am trying to implement PVN3D code on just the synthetic data (rendered data) which is produced from blender. Instead of real images I gave synthetic images only ( parameters and attributes extracted from pkl files which were generated in render folder). To my surprise am getting 0 improvement in training i.e. the loss value stop after creating single initial .pth file in checkpoints folder. Can anyone tell why the training won't improve if i just use synthetic data or If the code itself is built to improve only when it gets some real images ? Any comment on this issue is welcome