StreamPETR
StreamPETR copied to clipboard
exiawsh
[ICCV 2023] StreamPETR: Exploring Object-Centric Temporal Modeling for Efficient Multi-View 3D Object Detection
您好，请问为什么是以Object-Centric的目标检测方法呢，文中将其与BEV时序以及perspective时序进行了比较，不过没有琢磨明白什么叫做以Object-Centric的目标检测呢？盼复，谢谢
[https://github.com/exiawsh/StreamPETR/blob/main/projects/mmdet3d_plugin/models/utils/positional_encoding.py#L25]([https://github.com/exiawsh/StreamPETR/blob/main/projects/mmdet3d_plugin/models/utils/positional_encoding.py#L25](https://github.com/exiawsh/StreamPETR/issues/url)) Is there any special design to transpose 'x' & 'y' dimentions ? ``` def pos2posemb3d(pos, num_pos_feats=128, temperature=10000): scale = 2 * math.pi dim_t = torch.arange(num_pos_feats//2, dtype=torch.float32, device=pos.device) dim_t =...
在推理过程中prev_exists始终是None，我理解应该表示前一帧始终不存在， x = data['prev_exists'] self.memory_timestamp = memory_refresh(self.memory_timestamp[:, :self.memory_len], x) self.memory_reference_point = memory_refresh(self.memory_reference_point[:, :self.memory_len], x) self.memory_embedding = memory_refresh(self.memory_embedding[:, :self.memory_len], x) self.memory_egopose = memory_refresh(self.memory_egopose[:, :self.memory_len], x) self.memory_velo = memory_refresh(self.memory_velo[:, :self.memory_len], x) 经过以上代码获取到的memory都是零矩阵，所以请问是否如我理解的这样推理时不使用历史帧？
Hello, I would like to ask for your help to change the current code to the code trained only with the current frame, where should I modify the current configuration?...
warnings.warn( /local/mnt2/workspace/vaisb/miniconda3/envs/vaisb_PyTorch_13/lib/python3.8/site-packages/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: libtorch_cuda_cu.so: cannot open shared object file: No such file or directory warn(f"Failed to load image Python extension: {e}") Traceback (most recent...
老师您好，我发现projects/configs/StreamPETR/stream_petr_vov_flash_800_bs2_seq_24e.py文件里用的是MultiHeadAttention,如下图所示：  但是projects/mmdet3d_plugin/models/utils/petr_transformer.py里面没有MultiHeadAttention，只有PETRMultiheadAttention，如下图：  但是代码仍然可以正常跑起来，但是在debug的时候，只能进入PETRMultiheadFlashAttention，没有进入MultiheadAttention（我debug几次，没有看到），如图：  于是我将projects/mmdet3d_plugin/models/utils/petr_transformer.py中的PETRTemporalTransformer进行实例化：我将projects/configs/StreamPETR/stream_petr_vov_flash_800_bs2_seq_24e.py中关于PETRTemporalTransformer的信息粘贴到了projects/mmdet3d_plugin/models/utils/petr_transformer.py中的class PETRTemporalTransformer(BaseModule)模块的结尾，如图：   仍然可以打印MultiheadAttention如图：  请问：1、 MultiHeadAttention和PETRMultiheadFlashAttention是同一个东西吗 2、 MultiHeadAttention和PETRMultiheadAttention在网络中分别起到什么作用 盼复！！！感谢老师
 您好，我在看论文的时候发现Propagation Transformer的交叉注意力中使用了两个图像特征，分别是来自t-1时刻和t时刻的图像特征，但是我在阅读源码的时候并没有发现t-1时刻的特征从哪里来的（或者是t时刻的特征），在Transformer中只是用了一个6层的解码器，并没有使用两个6层的解码器。请问是我哪里理解错了吗
Hi @exiawsh. Thanks for your great work. I am trying to train the model using the default config repdetr3d_eva02_800_bs2_seq_24e.py. This uses streaming training. How can I set the number of...
