Learning Camera Localization via Dense Scene Matching, CVPR2021

Hi, thanks for your interesting work.

But the link of image files of Cambridge landmarks seems to be invalid, can you share with me the image files you download?

I got an error when i run "python setup.py install". Any idea about the solution?

/home/msg/.local/lib/python3.6/site-packages/torch/include/ATen/core/TensorBody.h:303:30: note: declared here
   DeprecatedTypeProperties & type() const {
                              ^~~~
/usr/local/cuda/bin/nvcc -I/home/msg/.local/lib/python3.6/site-packages/torch/include -I/home/msg/.local/lib/python3.6/site-packages/torch/include/torch/csrc/api/include -I/home/msg/.local/lib/python3.6/site-packages/torch/include/TH -I/home/msg/.local/lib/python3.6/site-packages/torch/include/THC -I/usr/local/cuda/include -I/usr/local/include/python3.6m -c correlation/src/corr_proj_kernel.cu -o build/temp.linux-x86_64-3.6/correlation/src/corr_proj_kernel.o -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr --compiler-options '-fPIC' -O2 -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE="_gcc" -DPYBIND11_STDLIB="_libstdcpp" -DPYBIND11_BUILD_ABI="_cxxabi1011" -DTORCH_EXTENSION_NAME=correlation_proj -D_GLIBCXX_USE_CXX11_ABI=0 -gencode=arch=compute_52,code=compute_52 -gencode=arch=compute_52,code=sm_52 -std=c++14
correlation/src/corr_proj_kernel.cu(190): error: no instance of overloaded function "atomicAdd" matches the argument list
            argument types are: (double *, double)
          detected during instantiation of "void CorrelateDataBackward(at::PackedTensorAccessor32<scalar_t, 4UL, at::RestrictPtrTraits>, at::PackedTensorAccessor32<scalar_t, 4UL, at::RestrictPtrTraits>, at::PackedTensorAccessor32<scalar_t, 5UL, at::RestrictPtrTraits>, at::PackedTensorAccessor32<scalar_t, 4UL, at::RestrictPtrTraits>, at::PackedTensorAccessor32<scalar_t, 4UL, at::RestrictPtrTraits>, at::PackedTensorAccessor32<scalar_t, 4UL, at::RestrictPtrTraits>, at::PackedTensorAccessor32<scalar_t, 5UL, at::RestrictPtrTraits>, int, int, int, int) [with scalar_t=double]" 
(235): here

1 error detected in the compilation of "correlation/src/corr_proj_kernel.cu".
error: command '/usr/local/cuda/bin/nvcc' failed with exit status 1

Thanks for sharing your work.

Hello,

I try to run "python tools/video_test.py --config configs/7scenes.yaml" in Colab and I get the following error. Which Cuda version did you use? I could not find a solution.

Traceback (most recent call last):
  File "tools/video_test.py", line 14, in <module>
    from engine.launcher import *
  File "libs/engine/launcher.py", line 6, in <module>
    from model.arch.DSMNet import dsm_net
  File "libs/model/arch/DSMNet.py", line 5, in <module>
    from ..ops.correlation.modules.corr import (
  File "libs/model/ops/correlation/modules/corr.py", line 4, in <module>
    from ..functions.corr import correlation_op, correlation_proj_op
  File "libs/model/ops/correlation/functions/corr.py", line 3, in <module>
    import correlation_cuda
ModuleNotFoundError: No module named 'correlation_cuda'

Thanks for sharing your great work.

Thanks for sharing your work. I am trying to repeat your work but I have some questions about the training process:

1. I wonder whether you use the whole scannet dataset.
2. Since you offer the label file of scannet but the image file is not supplied, I want to confirm whether your training data is downloaded directly from the scannet's author without post-processing.

Thanks a lot.

Thanks for sharing your work. I try to use DSM on my own dataset. I have rgb image, depth image and corresponding camera pose for them. But I noticed that the program loaded the "deep_retrieve_full.bin" file and got a "topk" attribute for every frame. I have two questions about that.

1. Is the "topk" representing the sorted correlation values that described in the paragraph "Cost volume", chapter 3.3.1 in the paper?
2. How can I generate the "topk" information for my own dataset?

Thanks a lot.

Thanks for sharing your work. I try to use DSM on 7scenes datasets but I noticed that the camera extrinsic is loaded from "deep_retrieve_full.bin" and the translation vector is different from corresponding frame-xxxxxx.pose.txt. Why is there this difference?

I am sorry to ask you for help. For some urgent reasons I need the Cambridge Landmarks datasets but the official website seemed to be blocked and there's no other data source I could find. I'll appreciate much if you have another source to help me get the dataset!

https://github.com/Tangshitao/Dense-Scene-Matching/blob/4957fa3f41419c31a60ffc82e234f23b8050583f/libs/engine/launcher.py#L27 should the "self.cfg.TRAIN.train_iters " be changed to "int(self.cfg.TRAIN.iters/self.cfg.TRAIN.model_save_iters)"? otherwise，there are too many cycles.we will save 1600000 models?